@oneciel-ai/claude-any 0.1.65 → 0.1.66
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +603 -594
- package/claude-any-menu.py +1948 -1948
- package/claude_any.py +7360 -7285
- package/docs/README.ja.md +515 -507
- package/docs/README.ko.md +514 -506
- package/docs/README.zh.md +493 -485
- package/docs/manual.md +548 -548
- package/package.json +42 -42
package/README.md
CHANGED
|
@@ -5,24 +5,24 @@
|
|
|
5
5
|
</p>
|
|
6
6
|
|
|
7
7
|
|
|
8
|
-
|
|
9
|
-
| English | [한국어](docs/README.ko.md) | [日本語](docs/README.ja.md) | [中文](docs/README.zh.md) |
|
|
10
|
-
| --- | --- | --- | --- |
|
|
11
|
-
|
|
12
|
-
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
13
|
-
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
14
|
-
[](LICENSE)
|
|
15
|
-
|
|
16
|
-
> ## 🚀 Use the full Claude Code experience with free or low-cost LLMs
|
|
17
|
-
>
|
|
18
|
-
> - **Free** — [NVIDIA hosted NIM](https://build.nvidia.com/) (qwen3-coder-480b, gpt-oss, and friends) through the API Catalog.
|
|
19
|
-
> - **Low-cost** — [Ollama Cloud](https://ollama.com/cloud) for GLM, Qwen, DeepSeek, and other open-weight models at a fraction of frontier-model pricing.
|
|
20
|
-
> - **Free + local** — [Ollama](https://ollama.com/) or [vLLM](https://github.com/vllm-project/vllm) on your own GPU, fully offline.
|
|
21
|
-
> - **Plan Mode + Advisor ready** — Claude Any preserves Claude Code Plan Mode on non-Anthropic providers and adds an optional long-context Advisor model for review.
|
|
22
|
-
> - **Smooth free-model pacing** — Claude Code spends time reading files and running tools, and Claude Any uses that natural gap for RPM pacing so NVIDIA hosted free models feel usable even with strict per-minute limits.
|
|
23
|
-
>
|
|
24
|
-
> Provider, model, base URL, API key, streaming behavior, and LLM options are all selected from a console menu **before** Claude Code starts. Claude Code itself runs untouched with all of its native tooling, slash commands, and workflows.
|
|
25
|
-
|
|
8
|
+
|
|
9
|
+
| English | [한국어](docs/README.ko.md) | [日本語](docs/README.ja.md) | [中文](docs/README.zh.md) |
|
|
10
|
+
| --- | --- | --- | --- |
|
|
11
|
+
|
|
12
|
+
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
13
|
+
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
14
|
+
[](LICENSE)
|
|
15
|
+
|
|
16
|
+
> ## 🚀 Use the full Claude Code experience with free or low-cost LLMs
|
|
17
|
+
>
|
|
18
|
+
> - **Free** — [NVIDIA hosted NIM](https://build.nvidia.com/) (qwen3-coder-480b, gpt-oss, and friends) through the API Catalog.
|
|
19
|
+
> - **Low-cost** — [Ollama Cloud](https://ollama.com/cloud) for GLM, Qwen, DeepSeek, and other open-weight models at a fraction of frontier-model pricing.
|
|
20
|
+
> - **Free + local** — [Ollama](https://ollama.com/) or [vLLM](https://github.com/vllm-project/vllm) on your own GPU, fully offline.
|
|
21
|
+
> - **Plan Mode + Advisor ready** — Claude Any preserves Claude Code Plan Mode on non-Anthropic providers and adds an optional long-context Advisor model for review.
|
|
22
|
+
> - **Smooth free-model pacing** — Claude Code spends time reading files and running tools, and Claude Any uses that natural gap for RPM pacing so NVIDIA hosted free models feel usable even with strict per-minute limits.
|
|
23
|
+
>
|
|
24
|
+
> Provider, model, base URL, API key, streaming behavior, and LLM options are all selected from a console menu **before** Claude Code starts. Claude Code itself runs untouched with all of its native tooling, slash commands, and workflows.
|
|
25
|
+
|
|
26
26
|
## Today's Top 3 Benefits
|
|
27
27
|
|
|
28
28
|
### 2026-05-14
|
|
@@ -36,99 +36,99 @@
|
|
|
36
36
|
1. **Plan Mode works on non-Anthropic models** — Claude Any keeps Claude Code's Plan Mode usable even when the upstream provider is NVIDIA hosted, Ollama Cloud, local Ollama, vLLM, or NIM.
|
|
37
37
|
2. **Advisor review with a bigger model** — pick a long-context Advisor Model at launch, then use `/advisor` inside Claude Code to review the current task, blockers, and next concrete action.
|
|
38
38
|
3. **Free-model RPM limits feel smoother** — router-side RPM pacing uses the natural time spent reading files and running tools, so NVIDIA hosted free models can stay within per-minute limits with less visible waiting.
|
|
39
|
-
|
|
40
|
-
### Demo
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
NVIDIA hosted NIM (deepseek-4-flash) driving Claude Code through the claude-any router. [full mp4 ⤓](https://github.com/OneCielAI/claude-any/raw/main/demo/claude-any-nvidia-nim.mp4)
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
Ollama Cloud (glm-5.1) streamed through the claude-any router with SSE word-boundary chunking enabled. [full mp4 ⤓](https://github.com/OneCielAI/claude-any/raw/main/demo/claude-any-ollama-cloud.mp4)
|
|
49
|
-
|
|
50
|
-
---
|
|
51
|
-
|
|
52
|
-
Claude Any is a provider selector and compatibility launcher for Claude Code.
|
|
53
|
-
It lets you choose Anthropic, Ollama, Ollama Cloud, vLLM, NVIDIA hosted models,
|
|
54
|
-
or self-hosted NIM before Claude Code starts, then passes normal Claude Code
|
|
55
|
-
arguments through unchanged.
|
|
56
|
-
|
|
57
|
-
Credits: One Ciel LLC
|
|
58
|
-
|
|
59
|
-
Current version: `0.1.
|
|
60
|
-
|
|
61
|
-
## Why This Exists
|
|
62
|
-
|
|
63
|
-
Claude Any started from a practical need: even on the highest Claude Code plan,
|
|
64
|
-
long sessions can run out of available tokens or become blocked while waiting
|
|
65
|
-
for the next quota window. The goal is not to replace Claude Code, but to keep
|
|
66
|
-
work moving. Slower but usable providers such as NVIDIA NIM, Ollama Cloud,
|
|
67
|
-
vLLM, and local Ollama can act as hybrid third-party agents for summaries,
|
|
68
|
-
research, journaling, simple coding tasks, and delegated background work.
|
|
69
|
-
|
|
70
|
-
Another design goal is to keep as much of Claude Code's native experience as
|
|
71
|
-
possible. When a provider exposes an Anthropic-compatible endpoint, Claude Any
|
|
72
|
-
prefers that path so Claude Code tooling, permissions, model selection, and
|
|
73
|
-
workflow behavior remain close to the original. For capabilities that remote
|
|
74
|
-
providers cannot supply directly, such as web search, Claude Any adds separate
|
|
75
|
-
MCP-based tooling.
|
|
76
|
-
|
|
77
|
-
The pre-launch menu is console-first. Provider, model, base URL, API key, and
|
|
78
|
-
options are meant to be easy to review and change before Claude Code starts,
|
|
79
|
-
including over SSH.
|
|
80
|
-
|
|
81
|
-
macOS has not been fully tested by the maintainer yet, but Claude Any uses
|
|
82
|
-
portable Python and shell wrappers. If you hit a macOS issue, please report it.
|
|
83
|
-
|
|
84
|
-
- D. Yun
|
|
85
|
-
|
|
86
|
-
## Install
|
|
87
|
-
|
|
88
|
-
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
89
|
-
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
90
|
-
|
|
91
|
-
Requirements:
|
|
92
|
-
|
|
93
|
-
- Python 3.10+
|
|
94
|
-
- Claude Code installed and available as `claude`
|
|
95
|
-
- Node/npm (used for the install shim and optional MCP web tooling)
|
|
96
|
-
|
|
97
|
-
**Install from the npm registry (recommended):**
|
|
98
|
-
|
|
99
|
-
```sh
|
|
100
|
-
npm install -g @oneciel-ai/claude-any
|
|
101
|
-
```
|
|
102
|
-
|
|
103
|
-
```sh
|
|
104
|
-
claude-any
|
|
105
|
-
```
|
|
106
|
-
|
|
107
|
-
## Run Claude Code Headlessly
|
|
108
|
-
|
|
109
|
-
Use headless mode when a script, SSH session, CI job, or parent agent needs to
|
|
110
|
-
launch Claude Code directly without opening the pre-launch menu. `claude-any`
|
|
111
|
-
consumes `--ca-*` options, starts the required local router, and passes the
|
|
112
|
-
remaining arguments to Claude Code.
|
|
113
|
-
|
|
114
|
-
```sh
|
|
115
|
-
claude-any --ca-provider nvidia-hosted --ca-model z-ai/glm-4.7
|
|
116
|
-
```
|
|
117
|
-
|
|
118
|
-
```sh
|
|
119
|
-
claude-any --ca-provider ollama-cloud --ca-model glm-5.1
|
|
120
|
-
```
|
|
121
|
-
|
|
122
|
-
```sh
|
|
123
|
-
claude-any --ca-provider ollama --ca-base-url http://127.0.0.1:11434 --ca-model qwen3-coder
|
|
124
|
-
```
|
|
125
|
-
|
|
126
|
-
Run one non-interactive Claude Code prompt:
|
|
127
|
-
|
|
128
|
-
```sh
|
|
129
|
-
claude-any --ca-provider nvidia-hosted --ca-model z-ai/glm-4.7 --ca-no-update-check -p "Reply with OK only." --output-format text
|
|
130
|
-
```
|
|
131
|
-
|
|
39
|
+
|
|
40
|
+
### Demo
|
|
41
|
+
|
|
42
|
+
|
|
43
|
+
|
|
44
|
+
NVIDIA hosted NIM (deepseek-4-flash) driving Claude Code through the claude-any router. [full mp4 ⤓](https://github.com/OneCielAI/claude-any/raw/main/demo/claude-any-nvidia-nim.mp4)
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
|
|
48
|
+
Ollama Cloud (glm-5.1) streamed through the claude-any router with SSE word-boundary chunking enabled. [full mp4 ⤓](https://github.com/OneCielAI/claude-any/raw/main/demo/claude-any-ollama-cloud.mp4)
|
|
49
|
+
|
|
50
|
+
---
|
|
51
|
+
|
|
52
|
+
Claude Any is a provider selector and compatibility launcher for Claude Code.
|
|
53
|
+
It lets you choose Anthropic, Ollama, Ollama Cloud, vLLM, NVIDIA hosted models,
|
|
54
|
+
or self-hosted NIM before Claude Code starts, then passes normal Claude Code
|
|
55
|
+
arguments through unchanged.
|
|
56
|
+
|
|
57
|
+
Credits: One Ciel LLC
|
|
58
|
+
|
|
59
|
+
Current version: `0.1.66`
|
|
60
|
+
|
|
61
|
+
## Why This Exists
|
|
62
|
+
|
|
63
|
+
Claude Any started from a practical need: even on the highest Claude Code plan,
|
|
64
|
+
long sessions can run out of available tokens or become blocked while waiting
|
|
65
|
+
for the next quota window. The goal is not to replace Claude Code, but to keep
|
|
66
|
+
work moving. Slower but usable providers such as NVIDIA NIM, Ollama Cloud,
|
|
67
|
+
vLLM, and local Ollama can act as hybrid third-party agents for summaries,
|
|
68
|
+
research, journaling, simple coding tasks, and delegated background work.
|
|
69
|
+
|
|
70
|
+
Another design goal is to keep as much of Claude Code's native experience as
|
|
71
|
+
possible. When a provider exposes an Anthropic-compatible endpoint, Claude Any
|
|
72
|
+
prefers that path so Claude Code tooling, permissions, model selection, and
|
|
73
|
+
workflow behavior remain close to the original. For capabilities that remote
|
|
74
|
+
providers cannot supply directly, such as web search, Claude Any adds separate
|
|
75
|
+
MCP-based tooling.
|
|
76
|
+
|
|
77
|
+
The pre-launch menu is console-first. Provider, model, base URL, API key, and
|
|
78
|
+
options are meant to be easy to review and change before Claude Code starts,
|
|
79
|
+
including over SSH.
|
|
80
|
+
|
|
81
|
+
macOS has not been fully tested by the maintainer yet, but Claude Any uses
|
|
82
|
+
portable Python and shell wrappers. If you hit a macOS issue, please report it.
|
|
83
|
+
|
|
84
|
+
- D. Yun
|
|
85
|
+
|
|
86
|
+
## Install
|
|
87
|
+
|
|
88
|
+
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
89
|
+
[](https://www.npmjs.com/package/@oneciel-ai/claude-any)
|
|
90
|
+
|
|
91
|
+
Requirements:
|
|
92
|
+
|
|
93
|
+
- Python 3.10+
|
|
94
|
+
- Claude Code installed and available as `claude`
|
|
95
|
+
- Node/npm (used for the install shim and optional MCP web tooling)
|
|
96
|
+
|
|
97
|
+
**Install from the npm registry (recommended):**
|
|
98
|
+
|
|
99
|
+
```sh
|
|
100
|
+
npm install -g @oneciel-ai/claude-any
|
|
101
|
+
```
|
|
102
|
+
|
|
103
|
+
```sh
|
|
104
|
+
claude-any
|
|
105
|
+
```
|
|
106
|
+
|
|
107
|
+
## Run Claude Code Headlessly
|
|
108
|
+
|
|
109
|
+
Use headless mode when a script, SSH session, CI job, or parent agent needs to
|
|
110
|
+
launch Claude Code directly without opening the pre-launch menu. `claude-any`
|
|
111
|
+
consumes `--ca-*` options, starts the required local router, and passes the
|
|
112
|
+
remaining arguments to Claude Code.
|
|
113
|
+
|
|
114
|
+
```sh
|
|
115
|
+
claude-any --ca-provider nvidia-hosted --ca-model z-ai/glm-4.7
|
|
116
|
+
```
|
|
117
|
+
|
|
118
|
+
```sh
|
|
119
|
+
claude-any --ca-provider ollama-cloud --ca-model glm-5.1
|
|
120
|
+
```
|
|
121
|
+
|
|
122
|
+
```sh
|
|
123
|
+
claude-any --ca-provider ollama --ca-base-url http://127.0.0.1:11434 --ca-model qwen3-coder
|
|
124
|
+
```
|
|
125
|
+
|
|
126
|
+
Run one non-interactive Claude Code prompt:
|
|
127
|
+
|
|
128
|
+
```sh
|
|
129
|
+
claude-any --ca-provider nvidia-hosted --ca-model z-ai/glm-4.7 --ca-no-update-check -p "Reply with OK only." --output-format text
|
|
130
|
+
```
|
|
131
|
+
|
|
132
132
|
Use the saved provider/model and only skip the menu:
|
|
133
133
|
|
|
134
134
|
```sh
|
|
@@ -180,172 +180,172 @@ scripts because the secret does not appear in shell history.
|
|
|
180
180
|
|
|
181
181
|
More examples are in [Headless Examples](#headless-examples) and
|
|
182
182
|
[the full manual](docs/manual.md#headless-usage).
|
|
183
|
-
|
|
184
|
-
**Upgrade:**
|
|
185
|
-
|
|
186
|
-
```sh
|
|
187
|
-
npm update -g @oneciel-ai/claude-any
|
|
188
|
-
```
|
|
189
|
-
|
|
190
|
-
```sh
|
|
191
|
-
claude-any version
|
|
192
|
-
```
|
|
193
|
-
|
|
194
|
-
**Uninstall:**
|
|
195
|
-
|
|
196
|
-
```sh
|
|
197
|
-
npm uninstall -g @oneciel-ai/claude-any
|
|
198
|
-
```
|
|
199
|
-
|
|
200
|
-
### Alternative install paths
|
|
201
|
-
|
|
202
|
-
Install directly from the GitHub repository (useful for testing unreleased
|
|
203
|
-
commits between npm publishes):
|
|
204
|
-
|
|
205
|
-
```sh
|
|
206
|
-
npm install -g https://github.com/OneCielAI/claude-any.git
|
|
207
|
-
```
|
|
208
|
-
|
|
209
|
-
```sh
|
|
210
|
-
claude-any
|
|
211
|
-
```
|
|
212
|
-
|
|
213
|
-
POSIX source install:
|
|
214
|
-
|
|
215
|
-
```sh
|
|
216
|
-
git clone https://github.com/OneCielAI/claude-any.git
|
|
217
|
-
```
|
|
218
|
-
|
|
219
|
-
```sh
|
|
220
|
-
cd claude-any
|
|
221
|
-
```
|
|
222
|
-
|
|
223
|
-
```sh
|
|
224
|
-
./install.sh
|
|
225
|
-
```
|
|
226
|
-
|
|
227
|
-
```sh
|
|
228
|
-
claude-any
|
|
229
|
-
```
|
|
230
|
-
|
|
231
|
-
Windows PowerShell source install:
|
|
232
|
-
|
|
233
|
-
```powershell
|
|
234
|
-
git clone https://github.com/OneCielAI/claude-any.git
|
|
235
|
-
```
|
|
236
|
-
|
|
237
|
-
```powershell
|
|
238
|
-
cd claude-any
|
|
239
|
-
```
|
|
240
|
-
|
|
241
|
-
```powershell
|
|
242
|
-
.\install.ps1
|
|
243
|
-
```
|
|
244
|
-
|
|
245
|
-
```powershell
|
|
246
|
-
claude-any
|
|
247
|
-
```
|
|
248
|
-
|
|
249
|
-
### Releasing (maintainers)
|
|
250
|
-
|
|
251
|
-
The npm registry version is published automatically by the
|
|
252
|
-
[`Publish to npm`](.github/workflows/npm-publish.yml) workflow when a new
|
|
253
|
-
GitHub Release is published. The workflow needs an `NPM_TOKEN` repository
|
|
254
|
-
secret containing a granular access token for `@oneciel-ai/claude-any` with
|
|
255
|
-
*Bypass 2FA for publishing* enabled.
|
|
256
|
-
|
|
257
|
-
Release flow:
|
|
258
|
-
|
|
259
|
-
1. Bump `version` in `package.json` and `VERSION` in `claude_any.py`.
|
|
260
|
-
2. Add a Changelog entry.
|
|
261
|
-
3. `git tag -a vX.Y.Z -m "..." && git push origin vX.Y.Z`.
|
|
262
|
-
4. `gh release create vX.Y.Z --title "..." --notes "..."` — this triggers
|
|
263
|
-
the publish workflow.
|
|
264
|
-
|
|
265
|
-
|
|
266
|
-
|
|
267
|
-
|
|
268
|
-
## Demo
|
|
269
|
-
|
|
270
|
-
|
|
271
|
-
|
|
272
|
-
The demo sequence now shows provider selection, Base URL entry, model selection,
|
|
273
|
-
LLM options, and the compatibility test. The compatibility test checks a plain
|
|
274
|
-
text response, a required `tool_use`, and a `tool_result` follow-up before the
|
|
275
|
-
launcher recommends starting Claude Code.
|
|
276
|
-
|
|
277
|
-
Additional current screenshots:
|
|
278
|
-
|
|
279
|
-
| Provider | Base URL | Model | LLM options | Compatibility |
|
|
280
|
-
| --- | --- | --- | --- | --- |
|
|
281
|
-
|  |  |  |  |  |
|
|
282
|
-
|
|
283
|
-
See the [full manual](docs/manual.md) for provider setup, headless flags, and
|
|
284
|
-
troubleshooting. A downloadable demo video is available at
|
|
285
|
-
[docs/assets/claude-any-demo.mp4](docs/assets/claude-any-demo.mp4).
|
|
286
|
-
|
|
287
|
-
## Development Story
|
|
288
|
-
|
|
289
|
-
Claude Any was built through real integration tests: provider switching, model
|
|
290
|
-
discovery, API-key entry, compatibility tests, web-search tooling, timeout
|
|
291
|
-
handling, and native Claude Code behavior. The main lesson was that
|
|
292
|
-
Anthropic-compatible Messages endpoints are the cleanest integration path when a
|
|
293
|
-
provider supports them. Ollama, vLLM, and NIM can expose Anthropic-compatible
|
|
294
|
-
routes that preserve more of Claude Code's tooling model than a generic
|
|
295
|
-
OpenAI-compatible chat route.
|
|
296
|
-
|
|
297
|
-
Local inference was also tested with Qwen 3.6 27B Q4 through Ollama and vLLM on
|
|
298
|
-
RTX 5090 and MSI GB10-class hardware. It worked, but the speed should not be
|
|
299
|
-
judged against native Claude Code or Codex. In practice, some hosted/cloud
|
|
300
|
-
choices such as NVIDIA NIM and Ollama Cloud felt more useful for this hybrid
|
|
301
|
-
workflow than expected.
|
|
302
|
-
|
|
303
|
-
OpenAI-compatible endpoints were deliberately kept out of the primary path for
|
|
304
|
-
Claude Code use. In testing, tool-call translation through generic OpenAI chat
|
|
305
|
-
compatibility was more brittle around tool parameters, tool results, repeated
|
|
306
|
-
calls, retries, and model selection.
|
|
307
|
-
|
|
308
|
-
The most recent vLLM finding is that server-side tool-call parsing must match
|
|
309
|
-
the model family. For Claude Code, a vLLM server can be reachable and still fail
|
|
310
|
-
if `--tool-call-parser` is wrong. In particular, Qwen3-Coder should be served
|
|
311
|
-
with `--enable-auto-tool-choice --tool-call-parser qwen3_xml`; Hermes is for
|
|
312
|
-
Hermes-style models and some older Qwen tool templates. Claude Any now surfaces
|
|
313
|
-
this in the compatibility test instead of treating a simple text response as
|
|
314
|
-
enough.
|
|
315
|
-
|
|
316
|
-
## Recommended Uses
|
|
317
|
-
|
|
318
|
-
Claude Any is most useful where speed is less important than keeping background
|
|
319
|
-
work moving. Good fits include Docker host maintenance, Windows or Linux system
|
|
320
|
-
administration, cleanup scripts for unused files, periodic security checklists,
|
|
321
|
-
log review, Windows Event Log review, intrusion-attempt triage, and report
|
|
322
|
-
drafting.
|
|
323
|
-
|
|
324
|
-
It is not a replacement for dedicated security products, but it can help
|
|
325
|
-
administrators turn routine checks into repeatable scripts and readable reports.
|
|
326
|
-
It is useful for summarizing possible virus, ransomware, brute-force, or
|
|
327
|
-
remote-access intrusion attempts. In that sense, Claude Any can help you build a
|
|
328
|
-
free or low-cost system security watcher for routine checks, alerts, and
|
|
329
|
-
human-readable summaries.
|
|
330
|
-
|
|
331
|
-
For example, it can help turn requests such as "install PostgreSQL in a Docker
|
|
332
|
-
container" or "analyze today's Docker logs and email me a report" into concrete
|
|
333
|
-
commands, scripts, scheduled jobs, and summaries.
|
|
334
|
-
|
|
335
|
-
A practical pattern is tiered supervision: use smaller or cheaper models to
|
|
336
|
-
watch logs and detect possible issues, use a larger model to review findings,
|
|
337
|
-
write policy, and plan the response, then let smaller models execute routine
|
|
338
|
-
steps under that larger model's supervision.
|
|
339
|
-
|
|
340
|
-
## Features
|
|
341
|
-
|
|
342
|
-
- Pre-launch provider picker with English, Korean, Japanese, and Chinese UI.
|
|
343
|
-
- Provider-aware model list and custom model entry.
|
|
344
|
-
- API key entry outside the Claude Code chat input.
|
|
345
|
-
- LLM option presets for context window, output tokens, timeout, sampling, and
|
|
346
|
-
native compatibility.
|
|
347
|
-
- Compatibility test before launch, including text response, tool use, and
|
|
348
|
-
tool-result round trip checks.
|
|
183
|
+
|
|
184
|
+
**Upgrade:**
|
|
185
|
+
|
|
186
|
+
```sh
|
|
187
|
+
npm update -g @oneciel-ai/claude-any
|
|
188
|
+
```
|
|
189
|
+
|
|
190
|
+
```sh
|
|
191
|
+
claude-any version
|
|
192
|
+
```
|
|
193
|
+
|
|
194
|
+
**Uninstall:**
|
|
195
|
+
|
|
196
|
+
```sh
|
|
197
|
+
npm uninstall -g @oneciel-ai/claude-any
|
|
198
|
+
```
|
|
199
|
+
|
|
200
|
+
### Alternative install paths
|
|
201
|
+
|
|
202
|
+
Install directly from the GitHub repository (useful for testing unreleased
|
|
203
|
+
commits between npm publishes):
|
|
204
|
+
|
|
205
|
+
```sh
|
|
206
|
+
npm install -g https://github.com/OneCielAI/claude-any.git
|
|
207
|
+
```
|
|
208
|
+
|
|
209
|
+
```sh
|
|
210
|
+
claude-any
|
|
211
|
+
```
|
|
212
|
+
|
|
213
|
+
POSIX source install:
|
|
214
|
+
|
|
215
|
+
```sh
|
|
216
|
+
git clone https://github.com/OneCielAI/claude-any.git
|
|
217
|
+
```
|
|
218
|
+
|
|
219
|
+
```sh
|
|
220
|
+
cd claude-any
|
|
221
|
+
```
|
|
222
|
+
|
|
223
|
+
```sh
|
|
224
|
+
./install.sh
|
|
225
|
+
```
|
|
226
|
+
|
|
227
|
+
```sh
|
|
228
|
+
claude-any
|
|
229
|
+
```
|
|
230
|
+
|
|
231
|
+
Windows PowerShell source install:
|
|
232
|
+
|
|
233
|
+
```powershell
|
|
234
|
+
git clone https://github.com/OneCielAI/claude-any.git
|
|
235
|
+
```
|
|
236
|
+
|
|
237
|
+
```powershell
|
|
238
|
+
cd claude-any
|
|
239
|
+
```
|
|
240
|
+
|
|
241
|
+
```powershell
|
|
242
|
+
.\install.ps1
|
|
243
|
+
```
|
|
244
|
+
|
|
245
|
+
```powershell
|
|
246
|
+
claude-any
|
|
247
|
+
```
|
|
248
|
+
|
|
249
|
+
### Releasing (maintainers)
|
|
250
|
+
|
|
251
|
+
The npm registry version is published automatically by the
|
|
252
|
+
[`Publish to npm`](.github/workflows/npm-publish.yml) workflow when a new
|
|
253
|
+
GitHub Release is published. The workflow needs an `NPM_TOKEN` repository
|
|
254
|
+
secret containing a granular access token for `@oneciel-ai/claude-any` with
|
|
255
|
+
*Bypass 2FA for publishing* enabled.
|
|
256
|
+
|
|
257
|
+
Release flow:
|
|
258
|
+
|
|
259
|
+
1. Bump `version` in `package.json` and `VERSION` in `claude_any.py`.
|
|
260
|
+
2. Add a Changelog entry.
|
|
261
|
+
3. `git tag -a vX.Y.Z -m "..." && git push origin vX.Y.Z`.
|
|
262
|
+
4. `gh release create vX.Y.Z --title "..." --notes "..."` — this triggers
|
|
263
|
+
the publish workflow.
|
|
264
|
+
|
|
265
|
+
|
|
266
|
+
|
|
267
|
+
|
|
268
|
+
## Demo
|
|
269
|
+
|
|
270
|
+
|
|
271
|
+
|
|
272
|
+
The demo sequence now shows provider selection, Base URL entry, model selection,
|
|
273
|
+
LLM options, and the compatibility test. The compatibility test checks a plain
|
|
274
|
+
text response, a required `tool_use`, and a `tool_result` follow-up before the
|
|
275
|
+
launcher recommends starting Claude Code.
|
|
276
|
+
|
|
277
|
+
Additional current screenshots:
|
|
278
|
+
|
|
279
|
+
| Provider | Base URL | Model | LLM options | Compatibility |
|
|
280
|
+
| --- | --- | --- | --- | --- |
|
|
281
|
+
|  |  |  |  |  |
|
|
282
|
+
|
|
283
|
+
See the [full manual](docs/manual.md) for provider setup, headless flags, and
|
|
284
|
+
troubleshooting. A downloadable demo video is available at
|
|
285
|
+
[docs/assets/claude-any-demo.mp4](docs/assets/claude-any-demo.mp4).
|
|
286
|
+
|
|
287
|
+
## Development Story
|
|
288
|
+
|
|
289
|
+
Claude Any was built through real integration tests: provider switching, model
|
|
290
|
+
discovery, API-key entry, compatibility tests, web-search tooling, timeout
|
|
291
|
+
handling, and native Claude Code behavior. The main lesson was that
|
|
292
|
+
Anthropic-compatible Messages endpoints are the cleanest integration path when a
|
|
293
|
+
provider supports them. Ollama, vLLM, and NIM can expose Anthropic-compatible
|
|
294
|
+
routes that preserve more of Claude Code's tooling model than a generic
|
|
295
|
+
OpenAI-compatible chat route.
|
|
296
|
+
|
|
297
|
+
Local inference was also tested with Qwen 3.6 27B Q4 through Ollama and vLLM on
|
|
298
|
+
RTX 5090 and MSI GB10-class hardware. It worked, but the speed should not be
|
|
299
|
+
judged against native Claude Code or Codex. In practice, some hosted/cloud
|
|
300
|
+
choices such as NVIDIA NIM and Ollama Cloud felt more useful for this hybrid
|
|
301
|
+
workflow than expected.
|
|
302
|
+
|
|
303
|
+
OpenAI-compatible endpoints were deliberately kept out of the primary path for
|
|
304
|
+
Claude Code use. In testing, tool-call translation through generic OpenAI chat
|
|
305
|
+
compatibility was more brittle around tool parameters, tool results, repeated
|
|
306
|
+
calls, retries, and model selection.
|
|
307
|
+
|
|
308
|
+
The most recent vLLM finding is that server-side tool-call parsing must match
|
|
309
|
+
the model family. For Claude Code, a vLLM server can be reachable and still fail
|
|
310
|
+
if `--tool-call-parser` is wrong. In particular, Qwen3-Coder should be served
|
|
311
|
+
with `--enable-auto-tool-choice --tool-call-parser qwen3_xml`; Hermes is for
|
|
312
|
+
Hermes-style models and some older Qwen tool templates. Claude Any now surfaces
|
|
313
|
+
this in the compatibility test instead of treating a simple text response as
|
|
314
|
+
enough.
|
|
315
|
+
|
|
316
|
+
## Recommended Uses
|
|
317
|
+
|
|
318
|
+
Claude Any is most useful where speed is less important than keeping background
|
|
319
|
+
work moving. Good fits include Docker host maintenance, Windows or Linux system
|
|
320
|
+
administration, cleanup scripts for unused files, periodic security checklists,
|
|
321
|
+
log review, Windows Event Log review, intrusion-attempt triage, and report
|
|
322
|
+
drafting.
|
|
323
|
+
|
|
324
|
+
It is not a replacement for dedicated security products, but it can help
|
|
325
|
+
administrators turn routine checks into repeatable scripts and readable reports.
|
|
326
|
+
It is useful for summarizing possible virus, ransomware, brute-force, or
|
|
327
|
+
remote-access intrusion attempts. In that sense, Claude Any can help you build a
|
|
328
|
+
free or low-cost system security watcher for routine checks, alerts, and
|
|
329
|
+
human-readable summaries.
|
|
330
|
+
|
|
331
|
+
For example, it can help turn requests such as "install PostgreSQL in a Docker
|
|
332
|
+
container" or "analyze today's Docker logs and email me a report" into concrete
|
|
333
|
+
commands, scripts, scheduled jobs, and summaries.
|
|
334
|
+
|
|
335
|
+
A practical pattern is tiered supervision: use smaller or cheaper models to
|
|
336
|
+
watch logs and detect possible issues, use a larger model to review findings,
|
|
337
|
+
write policy, and plan the response, then let smaller models execute routine
|
|
338
|
+
steps under that larger model's supervision.
|
|
339
|
+
|
|
340
|
+
## Features
|
|
341
|
+
|
|
342
|
+
- Pre-launch provider picker with English, Korean, Japanese, and Chinese UI.
|
|
343
|
+
- Provider-aware model list and custom model entry.
|
|
344
|
+
- API key entry outside the Claude Code chat input.
|
|
345
|
+
- LLM option presets for context window, output tokens, timeout, sampling, and
|
|
346
|
+
native compatibility.
|
|
347
|
+
- Compatibility test before launch, including text response, tool use, and
|
|
348
|
+
tool-result round trip checks.
|
|
349
349
|
- Runtime context reporting for vLLM/NIM when `/v1/models` exposes
|
|
350
350
|
`max_model_len`.
|
|
351
351
|
- Ollama model-context catalog: `claude-any ollama-catalog` downloads
|
|
@@ -353,46 +353,55 @@ steps under that larger model's supervision.
|
|
|
353
353
|
context windows such as 256K Kimi and 1M DeepSeek models for preset filtering
|
|
354
354
|
and status display.
|
|
355
355
|
- Console-first pre-launch menu for SSH and terminal workflows.
|
|
356
|
-
- Native paths where providers expose Claude/Anthropic-compatible endpoints.
|
|
357
|
-
- Router mode for providers that need request/response adaptation.
|
|
358
|
-
- DuckDuckGo and fetch MCP wiring for non-native providers.
|
|
359
|
-
- Headless setup flags such as `--ca-provider`, `--ca-model`, `--ca-base-url`,
|
|
360
|
-
`--ca-api-key-env`, `--ca-ollama-option`, and `--ca-max-output-tokens`.
|
|
361
|
-
- Claude Code Plan Mode support on router-backed non-Anthropic providers,
|
|
362
|
-
including local handling for `EnterPlanMode` and plan artifacts.
|
|
363
|
-
- Optional `/advisor` slash command that routes the current task state to a
|
|
364
|
-
selected Advisor Model, useful for long-context review and next-step checks.
|
|
365
|
-
- Claude Code `statusLine` integration showing router RPM usage and wait time
|
|
366
|
-
in the bottom status area instead of polluting the chat transcript.
|
|
367
|
-
- Router-side RPM control for NVIDIA hosted, self-hosted NIM, Ollama, and
|
|
368
|
-
Ollama Cloud. `rate_limit_rpm=0` disables throttling while still showing the
|
|
369
|
-
last-60-seconds usage rate.
|
|
370
|
-
- Soft pacing subtracts time already spent reading files, running commands, and
|
|
371
|
-
waiting for tool results. In real coding sessions, those tool-call gaps absorb
|
|
372
|
-
much of the RPM spacing naturally, so providers such as NVIDIA hosted NIM can
|
|
373
|
-
stay within free-model limits without making every Claude Code turn feel
|
|
374
|
-
rate-limited.
|
|
375
|
-
- Streaming proxy for Ollama/Ollama Cloud router path — tokens are delivered
|
|
376
|
-
to Claude Code as they arrive instead of waiting for the full response.
|
|
377
|
-
- Per-provider `stream` on/off toggle and `stream_word_chunking` option to
|
|
378
|
-
batch text deltas at word boundaries, mitigating SSE fragmentation that can
|
|
379
|
-
break tool-call / JSON parsing in long streamed responses.
|
|
380
|
-
- LLM options menu shows the meaning of the highlighted row at the bottom of
|
|
381
|
-
the panel in the selected language (English, Korean, Japanese, Chinese), and
|
|
382
|
-
boolean rows (`Stream`, `Stream word chunking`, `Native compatibility`,
|
|
383
|
-
`Think`) toggle in place when you press Enter — no input prompt.
|
|
384
|
-
- Tool guard hook coverage extended to the full Claude Code hook surface,
|
|
385
|
-
including `WorktreeCreate` / `WorktreeRemove`, so non-git working directories
|
|
386
|
-
no longer fail Agent isolation with
|
|
387
|
-
`Cannot create agent worktree: not in a git repository...`.
|
|
388
|
-
- Config file caching — settings are read from disk once and reused until the
|
|
389
|
-
file changes, reducing per-request overhead in the router.
|
|
390
|
-
- Router control-plane endpoints for headless agent coordination:
|
|
391
|
-
`/ca/chat/messages`, `/ca/chat/wait`, `/ca/chat/stream`, `/ca/chat/files`,
|
|
392
|
-
and `/ca/plan/artifacts`.
|
|
393
|
-
|
|
356
|
+
- Native paths where providers expose Claude/Anthropic-compatible endpoints.
|
|
357
|
+
- Router mode for providers that need request/response adaptation.
|
|
358
|
+
- DuckDuckGo and fetch MCP wiring for non-native providers.
|
|
359
|
+
- Headless setup flags such as `--ca-provider`, `--ca-model`, `--ca-base-url`,
|
|
360
|
+
`--ca-api-key-env`, `--ca-ollama-option`, and `--ca-max-output-tokens`.
|
|
361
|
+
- Claude Code Plan Mode support on router-backed non-Anthropic providers,
|
|
362
|
+
including local handling for `EnterPlanMode` and plan artifacts.
|
|
363
|
+
- Optional `/advisor` slash command that routes the current task state to a
|
|
364
|
+
selected Advisor Model, useful for long-context review and next-step checks.
|
|
365
|
+
- Claude Code `statusLine` integration showing router RPM usage and wait time
|
|
366
|
+
in the bottom status area instead of polluting the chat transcript.
|
|
367
|
+
- Router-side RPM control for NVIDIA hosted, self-hosted NIM, Ollama, and
|
|
368
|
+
Ollama Cloud. `rate_limit_rpm=0` disables throttling while still showing the
|
|
369
|
+
last-60-seconds usage rate.
|
|
370
|
+
- Soft pacing subtracts time already spent reading files, running commands, and
|
|
371
|
+
waiting for tool results. In real coding sessions, those tool-call gaps absorb
|
|
372
|
+
much of the RPM spacing naturally, so providers such as NVIDIA hosted NIM can
|
|
373
|
+
stay within free-model limits without making every Claude Code turn feel
|
|
374
|
+
rate-limited.
|
|
375
|
+
- Streaming proxy for Ollama/Ollama Cloud router path — tokens are delivered
|
|
376
|
+
to Claude Code as they arrive instead of waiting for the full response.
|
|
377
|
+
- Per-provider `stream` on/off toggle and `stream_word_chunking` option to
|
|
378
|
+
batch text deltas at word boundaries, mitigating SSE fragmentation that can
|
|
379
|
+
break tool-call / JSON parsing in long streamed responses.
|
|
380
|
+
- LLM options menu shows the meaning of the highlighted row at the bottom of
|
|
381
|
+
the panel in the selected language (English, Korean, Japanese, Chinese), and
|
|
382
|
+
boolean rows (`Stream`, `Stream word chunking`, `Native compatibility`,
|
|
383
|
+
`Think`) toggle in place when you press Enter — no input prompt.
|
|
384
|
+
- Tool guard hook coverage extended to the full Claude Code hook surface,
|
|
385
|
+
including `WorktreeCreate` / `WorktreeRemove`, so non-git working directories
|
|
386
|
+
no longer fail Agent isolation with
|
|
387
|
+
`Cannot create agent worktree: not in a git repository...`.
|
|
388
|
+
- Config file caching — settings are read from disk once and reused until the
|
|
389
|
+
file changes, reducing per-request overhead in the router.
|
|
390
|
+
- Router control-plane endpoints for headless agent coordination:
|
|
391
|
+
`/ca/chat/messages`, `/ca/chat/wait`, `/ca/chat/stream`, `/ca/chat/files`,
|
|
392
|
+
and `/ca/plan/artifacts`.
|
|
393
|
+
|
|
394
394
|
## Changelog
|
|
395
395
|
|
|
396
|
+
### 0.1.66
|
|
397
|
+
|
|
398
|
+
- **TUI input fix for tmux/zsh**: portable menu prompts now handle visible text,
|
|
399
|
+
paste, Backspace/Delete, and Ctrl-U directly in raw terminal mode instead of
|
|
400
|
+
depending on fragile terminal echo restoration.
|
|
401
|
+
- **Issue-linked release flow**: npm publishing can now be triggered
|
|
402
|
+
automatically by pushing a `v*` tag, in addition to GitHub releases and manual
|
|
403
|
+
workflow dispatch.
|
|
404
|
+
|
|
396
405
|
### 0.1.65
|
|
397
406
|
|
|
398
407
|
- **Plan Mode unchanged-Read loop recovery**: router conversion now preserves the
|
|
@@ -579,282 +588,282 @@ steps under that larger model's supervision.
|
|
|
579
588
|
|
|
580
589
|
### 0.1.31
|
|
581
590
|
|
|
582
|
-
- **2-minute default upstream timeout**: existing saved longer bundled defaults are migrated to 120000 ms so gateway stalls fail faster.
|
|
583
|
-
- **Localized gateway retries**: 502/503/504 and socket timeout responses are
|
|
584
|
-
retried automatically, with retry progress shown in the selected UI language.
|
|
585
|
-
|
|
586
|
-
### 0.1.30
|
|
587
|
-
|
|
588
|
-
- **Headless launch docs moved up**: README now shows copy-ready examples for
|
|
589
|
-
launching Claude Code directly with `--ca-provider`, `--ca-model`, `-p`, and
|
|
590
|
-
`CLAUDE_ANY_SKIP_MENU=1` immediately after install.
|
|
591
|
-
- **NVIDIA hosted wording cleanup**: provider and lifecycle docs now describe
|
|
592
|
-
NVIDIA hosted as using the Claude Any local router, with no separate hosted
|
|
593
|
-
API Catalog proxy requirement.
|
|
594
|
-
|
|
595
|
-
### 0.1.29
|
|
596
|
-
|
|
597
|
-
- **NVIDIA compatibility test fix**: `claude-any test` now restarts the local
|
|
598
|
-
router before router-mode tests, so upgraded installs do not accidentally use
|
|
599
|
-
an old long-running router that still expects `nvd-claude-proxy`.
|
|
600
|
-
- **Clear NVIDIA router wording**: menu status now describes NVIDIA hosted as
|
|
601
|
-
using the local claude-any router instead of the retired local proxy path.
|
|
602
|
-
|
|
603
|
-
### 0.1.28
|
|
604
|
-
|
|
605
|
-
- **Plan Mode + Advisor headline**: document Claude Any's Plan Mode support for
|
|
606
|
-
router-backed non-Anthropic providers and the `/advisor` slash command backed
|
|
607
|
-
by a selected long-context Advisor Model.
|
|
608
|
-
- **Status-line RPM telemetry**: Claude Any installs a Claude Code `statusLine`
|
|
609
|
-
command that shows router RPM usage and the latest wait time in the bottom
|
|
610
|
-
status area, keeping rate-limit telemetry out of the chat transcript.
|
|
611
|
-
- **Soft RPM pacing for free hosted models**: NVIDIA hosted, self-hosted NIM,
|
|
612
|
-
Ollama, and Ollama Cloud can use router-side RPM pacing. The pacing subtracts
|
|
613
|
-
time already spent in file reads, command execution, and tool-result waits, so
|
|
614
|
-
normal coding tool-call gaps naturally absorb much of the RPM spacing.
|
|
615
|
-
- **Unlimited usage display**: `rate_limit_rpm=0` disables throttling while
|
|
616
|
-
still displaying the last-60-seconds request rate.
|
|
617
|
-
|
|
618
|
-
### 0.1.27
|
|
619
|
-
|
|
620
|
-
- **Plan mode support for non-Anthropic providers**: the router now keeps
|
|
621
|
-
`EnterPlanMode` available and supports Claude Code Plan mode even when the
|
|
622
|
-
upstream model does not reliably choose that internal tool. Forced
|
|
623
|
-
`tool_choice=EnterPlanMode` is answered locally with a valid Anthropic
|
|
624
|
-
`tool_use`, and long implementation requests that receive only a short or
|
|
625
|
-
empty non-actionable text response are promoted to `EnterPlanMode` using
|
|
626
|
-
language-agnostic structure checks.
|
|
627
|
-
- **Plan-mode self-tool handling**: unsupported Claude Code self-tools are
|
|
628
|
-
still stripped for non-Anthropic providers, but Plan-mode tools are handled
|
|
629
|
-
separately so planning can work instead of being disabled.
|
|
630
|
-
|
|
631
|
-
### 0.1.25
|
|
632
|
-
|
|
633
|
-
- **Plan-mode diagnostics**: set `~/.config/claude-any/log-level` to `TRACE`
|
|
634
|
-
to capture redacted request and response summaries in `requests.jsonl` /
|
|
635
|
-
`responses.jsonl`.
|
|
636
|
-
- **Headless agent chat service**: the router exposes a small HTTP control
|
|
637
|
-
plane for sub coding agents. Agents can post messages, poll updates after
|
|
638
|
-
the last seen message id, or wait on an SSE stream when they do not have
|
|
639
|
-
their own loop.
|
|
640
|
-
- **Plan artifact serving**: agents can create plan files through the router
|
|
641
|
-
and share stable local URLs, matching Claude Code's file/artifact-oriented
|
|
642
|
-
Plan-mode workflow without copying Anthropic's internal implementation.
|
|
643
|
-
|
|
644
|
-
### 0.1.24
|
|
645
|
-
|
|
646
|
-
- **First public npm release** under the correct scope: `@oneciel-ai/claude-any`. Earlier 0.1.x versions were never published to the registry; this is the version that is actually installable via `npm install -g @oneciel-ai/claude-any`.
|
|
647
|
-
|
|
648
|
-
### 0.1.23
|
|
649
|
-
|
|
650
|
-
- **Stream toggle**: each non-Anthropic provider now has a `stream_enabled`
|
|
651
|
-
knob in the LLM options menu, in `claude-anyctl ollama-options` /
|
|
652
|
-
`provider-options`, and in headless flags. When off, the router forces
|
|
653
|
-
`stream:false` upstream and returns the full response to Claude Code — a
|
|
654
|
-
workaround when streaming fragmentation breaks tool-call or JSON parsing.
|
|
655
|
-
- **Word-boundary streaming**: new `stream_word_chunking` option buffers SSE
|
|
656
|
-
text deltas to whitespace/word boundaries before flushing. Implemented for
|
|
657
|
-
both the Ollama router path and the native pass-through path (vLLM, NVIDIA
|
|
658
|
-
hosted, self-hosted NIM). Tool deltas and non-text SSE events pass through
|
|
659
|
-
unchanged.
|
|
660
|
-
- **All-hooks tool guard**: `install_tool_guard_hooks` now registers the full
|
|
661
|
-
set of Claude Code hook events (PreToolUse, PostToolUse, PostToolUseFailure,
|
|
662
|
-
PostToolBatch, PermissionRequest, PermissionDenied, SessionStart/End, Setup,
|
|
663
|
-
UserPromptSubmit/Expansion, Stop, StopFailure, InstructionsLoaded,
|
|
664
|
-
ConfigChange, CwdChanged, Notification, SubagentStart/Stop, TeammateIdle,
|
|
665
|
-
TaskCreated, TaskCompleted, PreCompact, PostCompact, WorktreeCreate,
|
|
666
|
-
WorktreeRemove, Elicitation, ElicitationResult). The WorktreeCreate handler
|
|
667
|
-
emits `worktreePath = base_path` so Agent isolation works in non-git
|
|
668
|
-
directories.
|
|
669
|
-
- **Windows hook compatibility**: `shell_command_string` now emits forward
|
|
670
|
-
slashes and POSIX quoting on Windows so Claude Code's sh-based hook runner
|
|
671
|
-
doesn't strip backslashes from paths like `C:\Users\...`.
|
|
672
|
-
- **LLM options UX**: per-row description footer rendered in the user's
|
|
673
|
-
selected language, and boolean toggles (`Stream`, `Stream word chunking`,
|
|
674
|
-
`Native compatibility`, `Think`) flip on Enter without a prompt.
|
|
675
|
-
|
|
676
|
-
### 0.1.22
|
|
677
|
-
|
|
678
|
-
- **Headless manual expansion**: expand the manual with practical headless setup, launch, testing, passthrough, and cleanup examples for automation and remote-server use.
|
|
679
|
-
|
|
680
|
-
### 0.1.21
|
|
681
|
-
|
|
682
|
-
- **Service lifecycle documentation**: clarify that Claude Any starts only the router/proxy services required for the selected provider at launch time, and `claude-any stop` is the explicit cleanup command.
|
|
683
|
-
|
|
684
|
-
### 0.1.20
|
|
685
|
-
|
|
686
|
-
- **NVIDIA hosted quick test**: `auto` mode now uses a text-only quick test for NVIDIA hosted providers, avoiding slow or flaky tool_use requests during menu checks. Use `smoke` for text + tool_use, or `full` for the full text/tool_use/tool_result round trip.
|
|
687
|
-
- **Menu test timeout**: the terminal menu now runs `claude-any test 60 auto`, which keeps the pre-launch test responsive for hosted models.
|
|
688
|
-
|
|
689
|
-
### 0.1.19
|
|
690
|
-
|
|
691
|
-
- **Faster compatibility tests**: `claude-any test` now supports `auto`, `smoke`, and `full` modes.
|
|
692
|
-
- **Menu default speedup**: the terminal menu runs `claude-any test 120 auto`, so NVIDIA hosted compatibility checks finish faster while full validation remains available with `claude-any test 180 full`.
|
|
693
|
-
|
|
694
|
-
### 0.1.18
|
|
695
|
-
|
|
696
|
-
- **NVIDIA hosted transient diagnostics**: compatibility tests now identify `RemoteDisconnected`, connection resets, and 502/503/504 responses from NVIDIA hosted backends as transient upstream/API Catalog failures.
|
|
697
|
-
- **NVIDIA proxy cleanup**: `claude-any stop` now also matches `nvd-claude-proxy` executable processes so stale proxy sessions are cleaned up reliably.
|
|
698
|
-
|
|
699
|
-
### 0.1.17
|
|
700
|
-
|
|
701
|
-
- **Menu compatibility-test timeout**: the terminal menu now runs compatibility tests with an explicit 180 s timeout and stops the child process if it exceeds the menu hard limit, so slow hosted models cannot leave the menu appearing indefinitely pending.
|
|
702
|
-
|
|
703
|
-
### 0.1.16
|
|
704
|
-
|
|
705
|
-
- **NVIDIA hosted proxy startup fix**: detect and launch an installed `nvd-claude-proxy`/`ncp` executable before falling back to `python -m nvd_claude_proxy.main`. This supports uv-tool installs where the proxy is available as a command but not importable from Claude Any's Python interpreter.
|
|
706
|
-
|
|
707
|
-
### 0.1.15
|
|
708
|
-
|
|
709
|
-
- **Ollama/Ollama Cloud tool-call streaming fix**: emit streamed tool calls using sequential Anthropic SSE content block indexes and `input_json_delta` payloads. This prevents Claude Code from rejecting malformed streamed tool-use blocks with `Invalid tool parameters`.
|
|
710
|
-
- **Tool guard auto-install**: non-Anthropic launches now merge the Claude Any tool guard into `~/.claude/settings.json` so generated tool inputs can be normalized before execution.
|
|
711
|
-
- **Tool-call diagnostics**: router-side tool calls are logged to `~/.config/claude-any/tool-calls.jsonl`, and Claude Code hook inputs are logged to `~/.claude/claude-any-tool-guard/tool-events.jsonl` for precise debugging.
|
|
712
|
-
- **Tool input normalization**: the guard now maps common aliases such as `path` to `file_path`, `cmd` to `command`, and `query` to `pattern`, and returns explicit guidance when required fields are missing.
|
|
713
|
-
|
|
714
|
-
### 0.1.14
|
|
715
|
-
|
|
716
|
-
- **SSH/terminal arrow-key compatibility**: rewrote `read_menu_key()` with a proper ANSI escape sequence parser and moved raw terminal setup into `portable_select()` so the terminal stays in raw mode for the entire menu loop. This prevents escape sequences from leaking to the screen when `ECHO` is restored between keystrokes. Arrow keys, Home, and End now work reliably in SSH sessions.
|
|
717
|
-
- **Test timeout**: default compatibility test timeout increased from 60 s to 120 s for slower cloud providers.
|
|
718
|
-
- **Ollama Cloud compatibility test fix**: added `"stream": false` to compatibility test requests so the router fetches a single JSON response from Ollama Cloud instead of SSE streaming, which was causing `post_json` to timeout while collecting all chunks.
|
|
719
|
-
|
|
720
|
-
### 0.1.13
|
|
721
|
-
|
|
722
|
-
- **Ollama streaming proxy**: The router now streams Ollama and Ollama Cloud
|
|
723
|
-
responses through to Claude Code in real time using Anthropic SSE format,
|
|
724
|
-
instead of buffering the entire response before delivery.
|
|
725
|
-
- **Config caching**: `load_config()` now caches the configuration file in
|
|
726
|
-
memory and only re-reads from disk when the file modification time changes.
|
|
727
|
-
This eliminates repeated file I/O and JSON parsing on every router request.
|
|
728
|
-
- **Token estimation caching**: `estimate_tokens()` now accepts an optional
|
|
729
|
-
cache dict to avoid redundant `json.dumps()` calls within a single request.
|
|
730
|
-
`ollama_chat_request` and `cap_output_tokens_for_context` share the same
|
|
731
|
-
cache when computing context window sizing.
|
|
732
|
-
|
|
733
|
-
### 0.1.12
|
|
734
|
-
|
|
735
|
-
- Refresh docs and demo assets.
|
|
736
|
-
|
|
737
|
-
### 0.1.11
|
|
738
|
-
|
|
739
|
-
- Validate tool call compatibility.
|
|
740
|
-
|
|
741
|
-
### 0.1.10
|
|
742
|
-
|
|
743
|
-
- Show runtime context in tests.
|
|
744
|
-
|
|
745
|
-
### 0.1.9
|
|
746
|
-
|
|
747
|
-
- Cap presets to server context.
|
|
748
|
-
|
|
749
|
-
### 0.1.8
|
|
750
|
-
|
|
751
|
-
- Localize LLM presets.
|
|
752
|
-
|
|
753
|
-
## Provider Notes
|
|
754
|
-
|
|
755
|
-
| Provider | Mode | Notes |
|
|
756
|
-
| --- | --- | --- |
|
|
757
|
-
| Anthropic | Native Claude Code | Uses Claude login or Anthropic API key. |
|
|
758
|
-
| Ollama | Native when available, router otherwise | Local Ollama normally needs no API key. Cloud models through local Ollama require `ollama signin` on the Ollama host. |
|
|
759
|
-
| Ollama Cloud | Router | Calls `https://ollama.com/api`; requires an Ollama API key. |
|
|
760
|
-
| vLLM | Native Anthropic-compatible endpoint | Use a vLLM endpoint that exposes Anthropic-compatible `/v1/messages`; match `--tool-call-parser` to the model family. |
|
|
761
|
-
| NVIDIA hosted | Router | Uses the NVIDIA hosted API Catalog through the Claude Any local router. |
|
|
762
|
-
| self-hosted NIM | Native Anthropic-compatible endpoint | Use the self-hosted NIM Anthropic-compatible endpoint. |
|
|
763
|
-
|
|
764
|
-
## Service Lifecycle
|
|
765
|
-
|
|
766
|
-
Claude Any does not keep every possible backend helper running all the time. The
|
|
767
|
-
normal lifecycle is:
|
|
768
|
-
|
|
769
|
-
- Before launch, managed router processes can be stopped with
|
|
770
|
-
`claude-any stop`.
|
|
771
|
-
- When `claude-any` starts Claude Code, it starts only the services required by
|
|
772
|
-
the selected provider.
|
|
773
|
-
- Ollama and Ollama Cloud router mode use the Claude Any router on
|
|
774
|
-
`127.0.0.1:8799`.
|
|
775
|
-
- NVIDIA hosted router mode uses the Claude Any router on `127.0.0.1:8799`;
|
|
776
|
-
hosted API Catalog models do not require a separate NVIDIA proxy.
|
|
777
|
-
- Run `claude-any stop` before a fresh provider-switch test if an old router is
|
|
778
|
-
still bound to the local port.
|
|
779
|
-
|
|
780
|
-
This keeps Claude Code pointed at one stable Claude Any entry point while still
|
|
781
|
-
letting provider-specific helpers start on demand.
|
|
782
|
-
|
|
783
|
-
For Qwen3-Coder on vLLM, start the server with a matching tool parser:
|
|
784
|
-
|
|
785
|
-
```sh
|
|
786
|
-
vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
|
|
787
|
-
--host 0.0.0.0 \
|
|
788
|
-
--port 8000 \
|
|
789
|
-
--served-model-name qwen3-coder-30b \
|
|
790
|
-
--max-model-len 65536 \
|
|
791
|
-
--enable-auto-tool-choice \
|
|
792
|
-
--tool-call-parser qwen3_xml
|
|
793
|
-
```
|
|
794
|
-
|
|
795
|
-
## Provider Links
|
|
796
|
-
|
|
797
|
-
- Ollama Cloud: [cloud overview](https://ollama.com/cloud), [API key settings](https://ollama.com/settings/keys), [authentication docs](https://docs.ollama.com/api/authentication).
|
|
798
|
-
- Ollama local Anthropic compatibility: [Ollama Anthropic API docs](https://docs.ollama.com/api/anthropic-compatibility).
|
|
799
|
-
- vLLM: [Claude Code integration](https://docs.vllm.ai/en/latest/serving/integrations/claude_code/), [tool calling](https://docs.vllm.ai/en/stable/features/tool_calling/), [project GitHub](https://github.com/vllm-project/vllm).
|
|
800
|
-
- NVIDIA hosted NIM: [NVIDIA API Catalog](https://build.nvidia.com/), [API Catalog quickstart](https://docs.api.nvidia.com/nim/docs/api-quickstart).
|
|
801
|
-
- Self-hosted NVIDIA NIM: [Claude Code with NIM](https://docs.nvidia.com/nim/large-language-models/latest/ai-assistant-integrations/claude-code.html), [NIM for LLMs getting started](https://docs.nvidia.com/nim/large-language-models/1.14.0/getting-started.html), [NGC personal keys](https://org.ngc.nvidia.com/setup/personal-keys).
|
|
802
|
-
|
|
803
|
-
## Headless Examples
|
|
804
|
-
|
|
805
|
-
Headless commands skip the pre-launch menu and launch Claude Code immediately.
|
|
806
|
-
Claude Any consumes `--ca-*` setup flags, starts the required router
|
|
807
|
-
services, then passes the remaining arguments to Claude Code.
|
|
808
|
-
|
|
809
|
-
```sh
|
|
810
|
-
claude-any --ca-provider ollama-cloud --ca-model glm-5.1
|
|
811
|
-
claude-any --ca-provider ollama --ca-base-url http://127.0.0.1:11434 --ca-model qwen3-coder
|
|
812
|
-
claude-any --ca-provider ollama-cloud --ca-api-key-env OLLAMA_API_KEY --ca-model qwen3-coder:480b:cloud
|
|
813
|
-
claude-any --ca-provider vllm --ca-base-url http://127.0.0.1:8000 --ca-model Qwen/Qwen3-Coder
|
|
814
|
-
claude-any --ca-no-update-check -p "Reply with OK only." --output-format text
|
|
815
|
-
```
|
|
816
|
-
|
|
817
|
-
All other arguments are passed through to Claude Code.
|
|
818
|
-
|
|
819
|
-
## Headless Agent Chat
|
|
820
|
-
|
|
821
|
-
When the claude-any router is running, sub agents can coordinate through local
|
|
822
|
-
HTTP endpoints without opening the menu:
|
|
823
|
-
|
|
824
|
-
```sh
|
|
825
|
-
# Send a message to a channel.
|
|
826
|
-
curl -s http://127.0.0.1:8799/ca/chat/messages \
|
|
827
|
-
-H 'content-type: application/json' \
|
|
828
|
-
-d '{"channel":"agents","sender_id":"codex","recipients":["kimi"],"message":"Need logs after id 42"}'
|
|
829
|
-
|
|
830
|
-
# Poll updates after the last message id.
|
|
831
|
-
curl -s 'http://127.0.0.1:8799/ca/chat/messages?channel=agents&recipient=kimi&after=42'
|
|
832
|
-
|
|
833
|
-
# Wait on a stream until messages arrive.
|
|
834
|
-
curl -N 'http://127.0.0.1:8799/ca/chat/stream?channel=agents&recipient=kimi&after=42&timeout=300'
|
|
835
|
-
|
|
836
|
-
# Publish a plan file and get a served URL.
|
|
837
|
-
curl -s http://127.0.0.1:8799/ca/plan/artifacts \
|
|
838
|
-
-H 'content-type: application/json' \
|
|
839
|
-
-d '{"title":"handoff","name":"handoff.md","content":"# Plan\n- step 1"}'
|
|
840
|
-
```
|
|
841
|
-
|
|
842
|
-
## Security
|
|
843
|
-
|
|
844
|
-
Do not commit runtime configuration or API keys. Claude Any stores local runtime
|
|
845
|
-
configuration under `~/.config/claude-any/`. NVIDIA hosted credentials used by
|
|
846
|
-
the optional proxy are stored under `~/.config/nvd-claude-proxy/`.
|
|
847
|
-
|
|
848
|
-
This repository should contain source, documentation, and demo assets only.
|
|
849
|
-
|
|
850
|
-
## Development
|
|
851
|
-
|
|
852
|
-
```sh
|
|
853
|
-
python -m py_compile claude_any.py claude-any-menu.py claude-any-tool-guard.py
|
|
854
|
-
python -m ruff check .
|
|
855
|
-
python scripts/make_demo_assets.py
|
|
856
|
-
```
|
|
857
|
-
|
|
858
|
-
## License
|
|
859
|
-
|
|
860
|
-
MIT. See [LICENSE](LICENSE).
|
|
591
|
+
- **2-minute default upstream timeout**: existing saved longer bundled defaults are migrated to 120000 ms so gateway stalls fail faster.
|
|
592
|
+
- **Localized gateway retries**: 502/503/504 and socket timeout responses are
|
|
593
|
+
retried automatically, with retry progress shown in the selected UI language.
|
|
594
|
+
|
|
595
|
+
### 0.1.30
|
|
596
|
+
|
|
597
|
+
- **Headless launch docs moved up**: README now shows copy-ready examples for
|
|
598
|
+
launching Claude Code directly with `--ca-provider`, `--ca-model`, `-p`, and
|
|
599
|
+
`CLAUDE_ANY_SKIP_MENU=1` immediately after install.
|
|
600
|
+
- **NVIDIA hosted wording cleanup**: provider and lifecycle docs now describe
|
|
601
|
+
NVIDIA hosted as using the Claude Any local router, with no separate hosted
|
|
602
|
+
API Catalog proxy requirement.
|
|
603
|
+
|
|
604
|
+
### 0.1.29
|
|
605
|
+
|
|
606
|
+
- **NVIDIA compatibility test fix**: `claude-any test` now restarts the local
|
|
607
|
+
router before router-mode tests, so upgraded installs do not accidentally use
|
|
608
|
+
an old long-running router that still expects `nvd-claude-proxy`.
|
|
609
|
+
- **Clear NVIDIA router wording**: menu status now describes NVIDIA hosted as
|
|
610
|
+
using the local claude-any router instead of the retired local proxy path.
|
|
611
|
+
|
|
612
|
+
### 0.1.28
|
|
613
|
+
|
|
614
|
+
- **Plan Mode + Advisor headline**: document Claude Any's Plan Mode support for
|
|
615
|
+
router-backed non-Anthropic providers and the `/advisor` slash command backed
|
|
616
|
+
by a selected long-context Advisor Model.
|
|
617
|
+
- **Status-line RPM telemetry**: Claude Any installs a Claude Code `statusLine`
|
|
618
|
+
command that shows router RPM usage and the latest wait time in the bottom
|
|
619
|
+
status area, keeping rate-limit telemetry out of the chat transcript.
|
|
620
|
+
- **Soft RPM pacing for free hosted models**: NVIDIA hosted, self-hosted NIM,
|
|
621
|
+
Ollama, and Ollama Cloud can use router-side RPM pacing. The pacing subtracts
|
|
622
|
+
time already spent in file reads, command execution, and tool-result waits, so
|
|
623
|
+
normal coding tool-call gaps naturally absorb much of the RPM spacing.
|
|
624
|
+
- **Unlimited usage display**: `rate_limit_rpm=0` disables throttling while
|
|
625
|
+
still displaying the last-60-seconds request rate.
|
|
626
|
+
|
|
627
|
+
### 0.1.27
|
|
628
|
+
|
|
629
|
+
- **Plan mode support for non-Anthropic providers**: the router now keeps
|
|
630
|
+
`EnterPlanMode` available and supports Claude Code Plan mode even when the
|
|
631
|
+
upstream model does not reliably choose that internal tool. Forced
|
|
632
|
+
`tool_choice=EnterPlanMode` is answered locally with a valid Anthropic
|
|
633
|
+
`tool_use`, and long implementation requests that receive only a short or
|
|
634
|
+
empty non-actionable text response are promoted to `EnterPlanMode` using
|
|
635
|
+
language-agnostic structure checks.
|
|
636
|
+
- **Plan-mode self-tool handling**: unsupported Claude Code self-tools are
|
|
637
|
+
still stripped for non-Anthropic providers, but Plan-mode tools are handled
|
|
638
|
+
separately so planning can work instead of being disabled.
|
|
639
|
+
|
|
640
|
+
### 0.1.25
|
|
641
|
+
|
|
642
|
+
- **Plan-mode diagnostics**: set `~/.config/claude-any/log-level` to `TRACE`
|
|
643
|
+
to capture redacted request and response summaries in `requests.jsonl` /
|
|
644
|
+
`responses.jsonl`.
|
|
645
|
+
- **Headless agent chat service**: the router exposes a small HTTP control
|
|
646
|
+
plane for sub coding agents. Agents can post messages, poll updates after
|
|
647
|
+
the last seen message id, or wait on an SSE stream when they do not have
|
|
648
|
+
their own loop.
|
|
649
|
+
- **Plan artifact serving**: agents can create plan files through the router
|
|
650
|
+
and share stable local URLs, matching Claude Code's file/artifact-oriented
|
|
651
|
+
Plan-mode workflow without copying Anthropic's internal implementation.
|
|
652
|
+
|
|
653
|
+
### 0.1.24
|
|
654
|
+
|
|
655
|
+
- **First public npm release** under the correct scope: `@oneciel-ai/claude-any`. Earlier 0.1.x versions were never published to the registry; this is the version that is actually installable via `npm install -g @oneciel-ai/claude-any`.
|
|
656
|
+
|
|
657
|
+
### 0.1.23
|
|
658
|
+
|
|
659
|
+
- **Stream toggle**: each non-Anthropic provider now has a `stream_enabled`
|
|
660
|
+
knob in the LLM options menu, in `claude-anyctl ollama-options` /
|
|
661
|
+
`provider-options`, and in headless flags. When off, the router forces
|
|
662
|
+
`stream:false` upstream and returns the full response to Claude Code — a
|
|
663
|
+
workaround when streaming fragmentation breaks tool-call or JSON parsing.
|
|
664
|
+
- **Word-boundary streaming**: new `stream_word_chunking` option buffers SSE
|
|
665
|
+
text deltas to whitespace/word boundaries before flushing. Implemented for
|
|
666
|
+
both the Ollama router path and the native pass-through path (vLLM, NVIDIA
|
|
667
|
+
hosted, self-hosted NIM). Tool deltas and non-text SSE events pass through
|
|
668
|
+
unchanged.
|
|
669
|
+
- **All-hooks tool guard**: `install_tool_guard_hooks` now registers the full
|
|
670
|
+
set of Claude Code hook events (PreToolUse, PostToolUse, PostToolUseFailure,
|
|
671
|
+
PostToolBatch, PermissionRequest, PermissionDenied, SessionStart/End, Setup,
|
|
672
|
+
UserPromptSubmit/Expansion, Stop, StopFailure, InstructionsLoaded,
|
|
673
|
+
ConfigChange, CwdChanged, Notification, SubagentStart/Stop, TeammateIdle,
|
|
674
|
+
TaskCreated, TaskCompleted, PreCompact, PostCompact, WorktreeCreate,
|
|
675
|
+
WorktreeRemove, Elicitation, ElicitationResult). The WorktreeCreate handler
|
|
676
|
+
emits `worktreePath = base_path` so Agent isolation works in non-git
|
|
677
|
+
directories.
|
|
678
|
+
- **Windows hook compatibility**: `shell_command_string` now emits forward
|
|
679
|
+
slashes and POSIX quoting on Windows so Claude Code's sh-based hook runner
|
|
680
|
+
doesn't strip backslashes from paths like `C:\Users\...`.
|
|
681
|
+
- **LLM options UX**: per-row description footer rendered in the user's
|
|
682
|
+
selected language, and boolean toggles (`Stream`, `Stream word chunking`,
|
|
683
|
+
`Native compatibility`, `Think`) flip on Enter without a prompt.
|
|
684
|
+
|
|
685
|
+
### 0.1.22
|
|
686
|
+
|
|
687
|
+
- **Headless manual expansion**: expand the manual with practical headless setup, launch, testing, passthrough, and cleanup examples for automation and remote-server use.
|
|
688
|
+
|
|
689
|
+
### 0.1.21
|
|
690
|
+
|
|
691
|
+
- **Service lifecycle documentation**: clarify that Claude Any starts only the router/proxy services required for the selected provider at launch time, and `claude-any stop` is the explicit cleanup command.
|
|
692
|
+
|
|
693
|
+
### 0.1.20
|
|
694
|
+
|
|
695
|
+
- **NVIDIA hosted quick test**: `auto` mode now uses a text-only quick test for NVIDIA hosted providers, avoiding slow or flaky tool_use requests during menu checks. Use `smoke` for text + tool_use, or `full` for the full text/tool_use/tool_result round trip.
|
|
696
|
+
- **Menu test timeout**: the terminal menu now runs `claude-any test 60 auto`, which keeps the pre-launch test responsive for hosted models.
|
|
697
|
+
|
|
698
|
+
### 0.1.19
|
|
699
|
+
|
|
700
|
+
- **Faster compatibility tests**: `claude-any test` now supports `auto`, `smoke`, and `full` modes.
|
|
701
|
+
- **Menu default speedup**: the terminal menu runs `claude-any test 120 auto`, so NVIDIA hosted compatibility checks finish faster while full validation remains available with `claude-any test 180 full`.
|
|
702
|
+
|
|
703
|
+
### 0.1.18
|
|
704
|
+
|
|
705
|
+
- **NVIDIA hosted transient diagnostics**: compatibility tests now identify `RemoteDisconnected`, connection resets, and 502/503/504 responses from NVIDIA hosted backends as transient upstream/API Catalog failures.
|
|
706
|
+
- **NVIDIA proxy cleanup**: `claude-any stop` now also matches `nvd-claude-proxy` executable processes so stale proxy sessions are cleaned up reliably.
|
|
707
|
+
|
|
708
|
+
### 0.1.17
|
|
709
|
+
|
|
710
|
+
- **Menu compatibility-test timeout**: the terminal menu now runs compatibility tests with an explicit 180 s timeout and stops the child process if it exceeds the menu hard limit, so slow hosted models cannot leave the menu appearing indefinitely pending.
|
|
711
|
+
|
|
712
|
+
### 0.1.16
|
|
713
|
+
|
|
714
|
+
- **NVIDIA hosted proxy startup fix**: detect and launch an installed `nvd-claude-proxy`/`ncp` executable before falling back to `python -m nvd_claude_proxy.main`. This supports uv-tool installs where the proxy is available as a command but not importable from Claude Any's Python interpreter.
|
|
715
|
+
|
|
716
|
+
### 0.1.15
|
|
717
|
+
|
|
718
|
+
- **Ollama/Ollama Cloud tool-call streaming fix**: emit streamed tool calls using sequential Anthropic SSE content block indexes and `input_json_delta` payloads. This prevents Claude Code from rejecting malformed streamed tool-use blocks with `Invalid tool parameters`.
|
|
719
|
+
- **Tool guard auto-install**: non-Anthropic launches now merge the Claude Any tool guard into `~/.claude/settings.json` so generated tool inputs can be normalized before execution.
|
|
720
|
+
- **Tool-call diagnostics**: router-side tool calls are logged to `~/.config/claude-any/tool-calls.jsonl`, and Claude Code hook inputs are logged to `~/.claude/claude-any-tool-guard/tool-events.jsonl` for precise debugging.
|
|
721
|
+
- **Tool input normalization**: the guard now maps common aliases such as `path` to `file_path`, `cmd` to `command`, and `query` to `pattern`, and returns explicit guidance when required fields are missing.
|
|
722
|
+
|
|
723
|
+
### 0.1.14
|
|
724
|
+
|
|
725
|
+
- **SSH/terminal arrow-key compatibility**: rewrote `read_menu_key()` with a proper ANSI escape sequence parser and moved raw terminal setup into `portable_select()` so the terminal stays in raw mode for the entire menu loop. This prevents escape sequences from leaking to the screen when `ECHO` is restored between keystrokes. Arrow keys, Home, and End now work reliably in SSH sessions.
|
|
726
|
+
- **Test timeout**: default compatibility test timeout increased from 60 s to 120 s for slower cloud providers.
|
|
727
|
+
- **Ollama Cloud compatibility test fix**: added `"stream": false` to compatibility test requests so the router fetches a single JSON response from Ollama Cloud instead of SSE streaming, which was causing `post_json` to timeout while collecting all chunks.
|
|
728
|
+
|
|
729
|
+
### 0.1.13
|
|
730
|
+
|
|
731
|
+
- **Ollama streaming proxy**: The router now streams Ollama and Ollama Cloud
|
|
732
|
+
responses through to Claude Code in real time using Anthropic SSE format,
|
|
733
|
+
instead of buffering the entire response before delivery.
|
|
734
|
+
- **Config caching**: `load_config()` now caches the configuration file in
|
|
735
|
+
memory and only re-reads from disk when the file modification time changes.
|
|
736
|
+
This eliminates repeated file I/O and JSON parsing on every router request.
|
|
737
|
+
- **Token estimation caching**: `estimate_tokens()` now accepts an optional
|
|
738
|
+
cache dict to avoid redundant `json.dumps()` calls within a single request.
|
|
739
|
+
`ollama_chat_request` and `cap_output_tokens_for_context` share the same
|
|
740
|
+
cache when computing context window sizing.
|
|
741
|
+
|
|
742
|
+
### 0.1.12
|
|
743
|
+
|
|
744
|
+
- Refresh docs and demo assets.
|
|
745
|
+
|
|
746
|
+
### 0.1.11
|
|
747
|
+
|
|
748
|
+
- Validate tool call compatibility.
|
|
749
|
+
|
|
750
|
+
### 0.1.10
|
|
751
|
+
|
|
752
|
+
- Show runtime context in tests.
|
|
753
|
+
|
|
754
|
+
### 0.1.9
|
|
755
|
+
|
|
756
|
+
- Cap presets to server context.
|
|
757
|
+
|
|
758
|
+
### 0.1.8
|
|
759
|
+
|
|
760
|
+
- Localize LLM presets.
|
|
761
|
+
|
|
762
|
+
## Provider Notes
|
|
763
|
+
|
|
764
|
+
| Provider | Mode | Notes |
|
|
765
|
+
| --- | --- | --- |
|
|
766
|
+
| Anthropic | Native Claude Code | Uses Claude login or Anthropic API key. |
|
|
767
|
+
| Ollama | Native when available, router otherwise | Local Ollama normally needs no API key. Cloud models through local Ollama require `ollama signin` on the Ollama host. |
|
|
768
|
+
| Ollama Cloud | Router | Calls `https://ollama.com/api`; requires an Ollama API key. |
|
|
769
|
+
| vLLM | Native Anthropic-compatible endpoint | Use a vLLM endpoint that exposes Anthropic-compatible `/v1/messages`; match `--tool-call-parser` to the model family. |
|
|
770
|
+
| NVIDIA hosted | Router | Uses the NVIDIA hosted API Catalog through the Claude Any local router. |
|
|
771
|
+
| self-hosted NIM | Native Anthropic-compatible endpoint | Use the self-hosted NIM Anthropic-compatible endpoint. |
|
|
772
|
+
|
|
773
|
+
## Service Lifecycle
|
|
774
|
+
|
|
775
|
+
Claude Any does not keep every possible backend helper running all the time. The
|
|
776
|
+
normal lifecycle is:
|
|
777
|
+
|
|
778
|
+
- Before launch, managed router processes can be stopped with
|
|
779
|
+
`claude-any stop`.
|
|
780
|
+
- When `claude-any` starts Claude Code, it starts only the services required by
|
|
781
|
+
the selected provider.
|
|
782
|
+
- Ollama and Ollama Cloud router mode use the Claude Any router on
|
|
783
|
+
`127.0.0.1:8799`.
|
|
784
|
+
- NVIDIA hosted router mode uses the Claude Any router on `127.0.0.1:8799`;
|
|
785
|
+
hosted API Catalog models do not require a separate NVIDIA proxy.
|
|
786
|
+
- Run `claude-any stop` before a fresh provider-switch test if an old router is
|
|
787
|
+
still bound to the local port.
|
|
788
|
+
|
|
789
|
+
This keeps Claude Code pointed at one stable Claude Any entry point while still
|
|
790
|
+
letting provider-specific helpers start on demand.
|
|
791
|
+
|
|
792
|
+
For Qwen3-Coder on vLLM, start the server with a matching tool parser:
|
|
793
|
+
|
|
794
|
+
```sh
|
|
795
|
+
vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
|
|
796
|
+
--host 0.0.0.0 \
|
|
797
|
+
--port 8000 \
|
|
798
|
+
--served-model-name qwen3-coder-30b \
|
|
799
|
+
--max-model-len 65536 \
|
|
800
|
+
--enable-auto-tool-choice \
|
|
801
|
+
--tool-call-parser qwen3_xml
|
|
802
|
+
```
|
|
803
|
+
|
|
804
|
+
## Provider Links
|
|
805
|
+
|
|
806
|
+
- Ollama Cloud: [cloud overview](https://ollama.com/cloud), [API key settings](https://ollama.com/settings/keys), [authentication docs](https://docs.ollama.com/api/authentication).
|
|
807
|
+
- Ollama local Anthropic compatibility: [Ollama Anthropic API docs](https://docs.ollama.com/api/anthropic-compatibility).
|
|
808
|
+
- vLLM: [Claude Code integration](https://docs.vllm.ai/en/latest/serving/integrations/claude_code/), [tool calling](https://docs.vllm.ai/en/stable/features/tool_calling/), [project GitHub](https://github.com/vllm-project/vllm).
|
|
809
|
+
- NVIDIA hosted NIM: [NVIDIA API Catalog](https://build.nvidia.com/), [API Catalog quickstart](https://docs.api.nvidia.com/nim/docs/api-quickstart).
|
|
810
|
+
- Self-hosted NVIDIA NIM: [Claude Code with NIM](https://docs.nvidia.com/nim/large-language-models/latest/ai-assistant-integrations/claude-code.html), [NIM for LLMs getting started](https://docs.nvidia.com/nim/large-language-models/1.14.0/getting-started.html), [NGC personal keys](https://org.ngc.nvidia.com/setup/personal-keys).
|
|
811
|
+
|
|
812
|
+
## Headless Examples
|
|
813
|
+
|
|
814
|
+
Headless commands skip the pre-launch menu and launch Claude Code immediately.
|
|
815
|
+
Claude Any consumes `--ca-*` setup flags, starts the required router
|
|
816
|
+
services, then passes the remaining arguments to Claude Code.
|
|
817
|
+
|
|
818
|
+
```sh
|
|
819
|
+
claude-any --ca-provider ollama-cloud --ca-model glm-5.1
|
|
820
|
+
claude-any --ca-provider ollama --ca-base-url http://127.0.0.1:11434 --ca-model qwen3-coder
|
|
821
|
+
claude-any --ca-provider ollama-cloud --ca-api-key-env OLLAMA_API_KEY --ca-model qwen3-coder:480b:cloud
|
|
822
|
+
claude-any --ca-provider vllm --ca-base-url http://127.0.0.1:8000 --ca-model Qwen/Qwen3-Coder
|
|
823
|
+
claude-any --ca-no-update-check -p "Reply with OK only." --output-format text
|
|
824
|
+
```
|
|
825
|
+
|
|
826
|
+
All other arguments are passed through to Claude Code.
|
|
827
|
+
|
|
828
|
+
## Headless Agent Chat
|
|
829
|
+
|
|
830
|
+
When the claude-any router is running, sub agents can coordinate through local
|
|
831
|
+
HTTP endpoints without opening the menu:
|
|
832
|
+
|
|
833
|
+
```sh
|
|
834
|
+
# Send a message to a channel.
|
|
835
|
+
curl -s http://127.0.0.1:8799/ca/chat/messages \
|
|
836
|
+
-H 'content-type: application/json' \
|
|
837
|
+
-d '{"channel":"agents","sender_id":"codex","recipients":["kimi"],"message":"Need logs after id 42"}'
|
|
838
|
+
|
|
839
|
+
# Poll updates after the last message id.
|
|
840
|
+
curl -s 'http://127.0.0.1:8799/ca/chat/messages?channel=agents&recipient=kimi&after=42'
|
|
841
|
+
|
|
842
|
+
# Wait on a stream until messages arrive.
|
|
843
|
+
curl -N 'http://127.0.0.1:8799/ca/chat/stream?channel=agents&recipient=kimi&after=42&timeout=300'
|
|
844
|
+
|
|
845
|
+
# Publish a plan file and get a served URL.
|
|
846
|
+
curl -s http://127.0.0.1:8799/ca/plan/artifacts \
|
|
847
|
+
-H 'content-type: application/json' \
|
|
848
|
+
-d '{"title":"handoff","name":"handoff.md","content":"# Plan\n- step 1"}'
|
|
849
|
+
```
|
|
850
|
+
|
|
851
|
+
## Security
|
|
852
|
+
|
|
853
|
+
Do not commit runtime configuration or API keys. Claude Any stores local runtime
|
|
854
|
+
configuration under `~/.config/claude-any/`. NVIDIA hosted credentials used by
|
|
855
|
+
the optional proxy are stored under `~/.config/nvd-claude-proxy/`.
|
|
856
|
+
|
|
857
|
+
This repository should contain source, documentation, and demo assets only.
|
|
858
|
+
|
|
859
|
+
## Development
|
|
860
|
+
|
|
861
|
+
```sh
|
|
862
|
+
python -m py_compile claude_any.py claude-any-menu.py claude-any-tool-guard.py
|
|
863
|
+
python -m ruff check .
|
|
864
|
+
python scripts/make_demo_assets.py
|
|
865
|
+
```
|
|
866
|
+
|
|
867
|
+
## License
|
|
868
|
+
|
|
869
|
+
MIT. See [LICENSE](LICENSE).
|