claude-code-langchain 0.1.0__tar.gz
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.
- claude_code_langchain-0.1.0/LICENSE +21 -0
- claude_code_langchain-0.1.0/PKG-INFO +254 -0
- claude_code_langchain-0.1.0/README.md +229 -0
- claude_code_langchain-0.1.0/pyproject.toml +86 -0
- claude_code_langchain-0.1.0/pyproject.toml.orig +64 -0
- claude_code_langchain-0.1.0/src/claude_code_langchain/__init__.py +37 -0
- claude_code_langchain-0.1.0/src/claude_code_langchain/chat_models.py +759 -0
- claude_code_langchain-0.1.0/src/claude_code_langchain/py.typed +0 -0
- claude_code_langchain-0.1.0/src/claude_code_langchain/tools.py +42 -0
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Moésio Filho
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,254 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: claude-code-langchain
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: LangChain chat model for Claude Code, built on the Claude Agent SDK
|
|
5
|
+
Keywords: langchain,claude,anthropic,claude-code,claude-agent-sdk
|
|
6
|
+
Author: Moésio Filho
|
|
7
|
+
Author-email: Moésio Filho <moesio.f@gmail.com>
|
|
8
|
+
License-Expression: MIT
|
|
9
|
+
License-File: LICENSE
|
|
10
|
+
Classifier: Development Status :: 3 - Alpha
|
|
11
|
+
Classifier: Intended Audience :: Developers
|
|
12
|
+
Classifier: Programming Language :: Python :: 3
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
14
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
15
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
16
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
17
|
+
Classifier: Topic :: Software Development :: Libraries
|
|
18
|
+
Requires-Dist: claude-agent-sdk>=0.2.0
|
|
19
|
+
Requires-Dist: langchain-core>=1.0.0,<2.0.0
|
|
20
|
+
Requires-Python: >=3.10
|
|
21
|
+
Project-URL: Homepage, https://github.com/moesio-f/claude-code-langchain
|
|
22
|
+
Project-URL: Repository, https://github.com/moesio-f/claude-code-langchain
|
|
23
|
+
Project-URL: Issues, https://github.com/moesio-f/claude-code-langchain/issues
|
|
24
|
+
Description-Content-Type: text/markdown
|
|
25
|
+
|
|
26
|
+
# claude-code-langchain
|
|
27
|
+
|
|
28
|
+
[](https://pypi.org/project/claude-code-langchain/)
|
|
29
|
+
[](https://pypi.org/project/claude-code-langchain/)
|
|
30
|
+
[](https://github.com/moesio-f/claude-code-langchain/blob/main/LICENSE)
|
|
31
|
+
|
|
32
|
+
A LangChain chat model that runs Claude through the **Claude Code CLI**, using the
|
|
33
|
+
[Claude Agent SDK](https://github.com/anthropics/claude-agent-sdk-python).
|
|
34
|
+
|
|
35
|
+
> **Not affiliated with Anthropic.** This is an independent community project.
|
|
36
|
+
> "Anthropic", "Claude", and "Claude Code" are trademarks of Anthropic PBC.
|
|
37
|
+
|
|
38
|
+
Authentication is delegated entirely to the `claude` CLI, so the model runs with whatever
|
|
39
|
+
the CLI is already configured with. For products or services that serve other users,
|
|
40
|
+
Anthropic requires API-key authentication — see the
|
|
41
|
+
[Claude Code legal and compliance page](https://code.claude.com/docs/en/legal-and-compliance).
|
|
42
|
+
|
|
43
|
+
> Built on the ideas of [ChamaruAmasara/langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code),
|
|
44
|
+
> rewritten for the current stack: the new `claude-agent-sdk` (replacing the deprecated
|
|
45
|
+
> `claude-code-sdk`) and LangChain 1.0+.
|
|
46
|
+
|
|
47
|
+
## What the rewrite changes
|
|
48
|
+
|
|
49
|
+
| Area | Original (`claude-code-sdk`) | Here (`claude-agent-sdk`) |
|
|
50
|
+
|---|---|---|
|
|
51
|
+
| SDK options | `ClaudeCodeOptions` | `ClaudeAgentOptions` |
|
|
52
|
+
| Tool calling | Tool schemas injected into the system prompt, JSON parsed out of the reply | Real Anthropic tool use: tools are served over an in-process MCP server, so `AIMessage.tool_calls` comes from actual `tool_use` blocks |
|
|
53
|
+
| Structured output | Prompted, best-effort | The CLI's `--json-schema` output format, validated against your schema |
|
|
54
|
+
| Extended thinking | A "think step by step" hint appended to the prompt | The SDK's native `thinking` config, returned as a `reasoning` content block |
|
|
55
|
+
| Effort | `--effort` via `extra_args` | First-class `effort` option, `low` → `max` |
|
|
56
|
+
| Token usage | Not available | `usage_metadata`, including cache reads/writes, plus cost in `response_metadata` |
|
|
57
|
+
| Built-in tools | Implicitly on, gated by `max_turns` | Explicit `builtin_tools`; off by default, so plain chat stays plain |
|
|
58
|
+
| LangChain | 0.3 | 1.0+ (v1 content blocks, `usage_metadata`, standard `bind_tools` / `with_structured_output`) |
|
|
59
|
+
|
|
60
|
+
## Requirements
|
|
61
|
+
|
|
62
|
+
- Python 3.10+
|
|
63
|
+
- Node.js 18+ and the Claude Code CLI: `npm install -g @anthropic-ai/claude-code && claude`
|
|
64
|
+
- The CLI authenticated with the method you intend to use (API key, Claude subscription, or
|
|
65
|
+
a supported cloud provider)
|
|
66
|
+
|
|
67
|
+
```bash
|
|
68
|
+
pip install claude-code-langchain
|
|
69
|
+
```
|
|
70
|
+
|
|
71
|
+
## Quick start
|
|
72
|
+
|
|
73
|
+
```python
|
|
74
|
+
from claude_code_langchain import ChatClaudeCode
|
|
75
|
+
|
|
76
|
+
llm = ChatClaudeCode(model="claude-opus-5")
|
|
77
|
+
print(llm.invoke("What is the capital of France?").text)
|
|
78
|
+
```
|
|
79
|
+
|
|
80
|
+
`model` accepts a model id (`claude-opus-5`, `claude-sonnet-5`, `claude-haiku-4-5`) or a CLI
|
|
81
|
+
alias (`opus`, `sonnet`, `haiku`). Leave it unset to use whatever the CLI is configured with.
|
|
82
|
+
|
|
83
|
+
### Messages, streaming, async
|
|
84
|
+
|
|
85
|
+
```python
|
|
86
|
+
from langchain_core.messages import HumanMessage, SystemMessage
|
|
87
|
+
|
|
88
|
+
llm.invoke([SystemMessage("Be concise."), HumanMessage("Explain generators.")])
|
|
89
|
+
|
|
90
|
+
for chunk in llm.stream("Count from 1 to 5"):
|
|
91
|
+
print(chunk.text, end="", flush=True)
|
|
92
|
+
|
|
93
|
+
await llm.ainvoke("Hello") # also: astream, batch, abatch
|
|
94
|
+
```
|
|
95
|
+
|
|
96
|
+
### Tool calling
|
|
97
|
+
|
|
98
|
+
```python
|
|
99
|
+
from langchain_core.tools import tool
|
|
100
|
+
|
|
101
|
+
|
|
102
|
+
@tool
|
|
103
|
+
def get_weather(city: str) -> str:
|
|
104
|
+
"""Get the weather for a city."""
|
|
105
|
+
return "12C and sunny"
|
|
106
|
+
|
|
107
|
+
|
|
108
|
+
response = llm.bind_tools([get_weather]).invoke("What is the weather in Paris?")
|
|
109
|
+
response.tool_calls
|
|
110
|
+
# [{'name': 'get_weather', 'args': {'city': 'Paris'}, 'id': 'toolu_…', 'type': 'tool_call'}]
|
|
111
|
+
```
|
|
112
|
+
|
|
113
|
+
Your tools are exposed to the CLI over an in-process MCP server, but they are never executed
|
|
114
|
+
there — the call is handed back to you, exactly like any other LangChain chat model, so it
|
|
115
|
+
works with LangGraph's `create_agent` and the usual `ToolNode` loop.
|
|
116
|
+
|
|
117
|
+
### Structured output
|
|
118
|
+
|
|
119
|
+
```python
|
|
120
|
+
from pydantic import BaseModel
|
|
121
|
+
|
|
122
|
+
|
|
123
|
+
class Person(BaseModel):
|
|
124
|
+
"""A person."""
|
|
125
|
+
|
|
126
|
+
name: str
|
|
127
|
+
age: int
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
llm.with_structured_output(Person).invoke("Bob is 42.")
|
|
131
|
+
# Person(name='Bob', age=42)
|
|
132
|
+
```
|
|
133
|
+
|
|
134
|
+
This uses the CLI's native JSON-schema output mode, so the response is constrained rather
|
|
135
|
+
than merely requested. Pass `include_raw=True` to get `{"raw", "parsed", "parsing_error"}`.
|
|
136
|
+
|
|
137
|
+
### Thinking and effort
|
|
138
|
+
|
|
139
|
+
```python
|
|
140
|
+
llm = ChatClaudeCode(
|
|
141
|
+
model="claude-opus-5",
|
|
142
|
+
thinking={"type": "adaptive", "display": "summarized"},
|
|
143
|
+
effort="high",
|
|
144
|
+
)
|
|
145
|
+
message = llm.invoke("Prove that sqrt(2) is irrational.")
|
|
146
|
+
message.content # [{'type': 'reasoning', ...}, {'type': 'text', ...}]
|
|
147
|
+
message.text # just the answer
|
|
148
|
+
```
|
|
149
|
+
|
|
150
|
+
### Agentic mode
|
|
151
|
+
|
|
152
|
+
Claude Code's own tools are **off by default**, so the model behaves like a normal chat model.
|
|
153
|
+
Turn them on when you want an agent that can touch your filesystem:
|
|
154
|
+
|
|
155
|
+
```python
|
|
156
|
+
from claude_code_langchain import READ_ONLY_TOOLS, ChatClaudeCode, ClaudeTool
|
|
157
|
+
|
|
158
|
+
reader = ChatClaudeCode(
|
|
159
|
+
builtin_tools=READ_ONLY_TOOLS, # Read, Glob, Grep
|
|
160
|
+
allowed_tools=READ_ONLY_TOOLS, # run without asking for permission
|
|
161
|
+
max_turns=10,
|
|
162
|
+
cwd="/path/to/project",
|
|
163
|
+
)
|
|
164
|
+
reader.invoke("Find every TODO comment in this project")
|
|
165
|
+
|
|
166
|
+
full = ChatClaudeCode(
|
|
167
|
+
builtin_tools="claude_code", # the complete Claude Code toolset
|
|
168
|
+
permission_mode="acceptEdits",
|
|
169
|
+
max_turns=20,
|
|
170
|
+
cwd="/path/to/project",
|
|
171
|
+
)
|
|
172
|
+
```
|
|
173
|
+
|
|
174
|
+
Presets: `READ_ONLY_TOOLS`, `WRITE_TOOLS`, `SHELL_TOOLS`, `NETWORK_TOOLS`, `ALL_TOOLS`, plus the
|
|
175
|
+
`ClaudeTool` enum for individual names.
|
|
176
|
+
|
|
177
|
+
### Sessions
|
|
178
|
+
|
|
179
|
+
Every response carries the CLI session id, which you can resume later:
|
|
180
|
+
|
|
181
|
+
```python
|
|
182
|
+
first = llm.invoke("Remember the number 7.")
|
|
183
|
+
session = first.response_metadata["session_id"]
|
|
184
|
+
|
|
185
|
+
resumed = ChatClaudeCode(session_id=session).invoke("What number did I mention?")
|
|
186
|
+
```
|
|
187
|
+
|
|
188
|
+
## Configuration
|
|
189
|
+
|
|
190
|
+
| Option | Purpose |
|
|
191
|
+
|---|---|
|
|
192
|
+
| `model`, `system_prompt` | Model id/alias; system prompt used when no `SystemMessage` is present |
|
|
193
|
+
| `thinking`, `effort` | Extended thinking config; `low` / `medium` / `high` / `xhigh` / `max` |
|
|
194
|
+
| `max_turns`, `max_budget_usd` | Agent loop limits (defaults to a single turn) |
|
|
195
|
+
| `builtin_tools`, `allowed_tools`, `disallowed_tools`, `permission_mode` | Which Claude Code tools exist, which run unattended, and how permissions are handled |
|
|
196
|
+
| `mcp_servers` | Additional MCP servers for the session |
|
|
197
|
+
| `cwd`, `add_dirs`, `env`, `cli_path`, `settings`, `setting_sources` | Where and how the CLI subprocess runs |
|
|
198
|
+
| `session_id`, `fork_session` | Resume or branch a previous session |
|
|
199
|
+
| `extra_args` | Any other CLI flag |
|
|
200
|
+
|
|
201
|
+
`setting_sources` defaults to `[]` so a run is not affected by your local Claude settings or
|
|
202
|
+
`CLAUDE.md`. Pass `None` to load user, project and local settings like the interactive CLI.
|
|
203
|
+
|
|
204
|
+
## Compared with `ChatAnthropic`
|
|
205
|
+
|
|
206
|
+
| Feature | `ChatAnthropic` | `ChatClaudeCode` |
|
|
207
|
+
|---|---|---|
|
|
208
|
+
| `invoke` / `stream` / `batch` / async | ✅ | ✅ |
|
|
209
|
+
| System messages, image input | ✅ | ✅ |
|
|
210
|
+
| Tool calling, structured output | ✅ | ✅ |
|
|
211
|
+
| Extended thinking, effort | ✅ | ✅ |
|
|
212
|
+
| Token usage and cost | ✅ | ✅ (plus USD cost per call) |
|
|
213
|
+
| Authentication | API key | Delegated to the `claude` CLI |
|
|
214
|
+
| Built-in filesystem/bash/web tools, MCP servers, subagents, skills | ❌ | ✅ |
|
|
215
|
+
| `temperature`, `top_p`, `top_k`, `stop` | ✅ | Accepted for compatibility, ignored by the CLI |
|
|
216
|
+
| Multi-turn history | Sent verbatim | Flattened into one prompt (see below) |
|
|
217
|
+
|
|
218
|
+
### Known limitations
|
|
219
|
+
|
|
220
|
+
- The CLI owns conversation state, so a message list is flattened into a single prompt with
|
|
221
|
+
labelled turns. Prior tool calls and results are rendered as text. Use `session_id` when you
|
|
222
|
+
want the CLI's own memory of a conversation instead.
|
|
223
|
+
- Sampling parameters and stop sequences are not exposed by the CLI.
|
|
224
|
+
- Streaming yields text only; thinking arrives on the final non-streamed message.
|
|
225
|
+
- Each call spawns a `claude` subprocess, so latency is higher than a direct API call.
|
|
226
|
+
|
|
227
|
+
## Development
|
|
228
|
+
|
|
229
|
+
```bash
|
|
230
|
+
uv sync --group dev
|
|
231
|
+
uv run pytest # unit tests, no CLI needed
|
|
232
|
+
uv run ruff check . && uv run mypy src
|
|
233
|
+
CLAUDE_CODE_LANGCHAIN_INTEGRATION=1 uv run pytest tests/integration # runs the CLI
|
|
234
|
+
```
|
|
235
|
+
|
|
236
|
+
## Related projects
|
|
237
|
+
|
|
238
|
+
Other LangChain adapters for Claude Code:
|
|
239
|
+
|
|
240
|
+
| Project | Notes |
|
|
241
|
+
|---|---|
|
|
242
|
+
| [ChamaruAmasara/langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code) | The original project this one was inspired by; on PyPI as `langchain-claude-code-cli` |
|
|
243
|
+
| [thehumanworks/langchain-claude-code](https://github.com/thehumanworks/langchain-claude-code) | Wrapper for the Claude Agent SDK; on PyPI as `langchain-claude-code` |
|
|
244
|
+
| [brhodes/langchain-claude-agent-sdk](https://codeberg.org/brhodes/langchain-claude-agent-sdk) | `langchain-claude-agent-sdk` on PyPI (hosted on Codeberg rather than GitHub) |
|
|
245
|
+
| [langchain-ai/langchain#40131](https://github.com/langchain-ai/langchain/issues/40131) | Proposal to add an Agent SDK-backed model to `langchain-anthropic` |
|
|
246
|
+
|
|
247
|
+
Each takes a different approach to authentication and conversation state. Some accept OAuth
|
|
248
|
+
tokens directly; review the auth method and Anthropic's terms before using one in production.
|
|
249
|
+
|
|
250
|
+
## Credits
|
|
251
|
+
|
|
252
|
+
Original project: [langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code)
|
|
253
|
+
by Chamaru Amasara. This package is an independent rewrite on the Claude Agent SDK and
|
|
254
|
+
LangChain 1.0+.
|
|
@@ -0,0 +1,229 @@
|
|
|
1
|
+
# claude-code-langchain
|
|
2
|
+
|
|
3
|
+
[](https://pypi.org/project/claude-code-langchain/)
|
|
4
|
+
[](https://pypi.org/project/claude-code-langchain/)
|
|
5
|
+
[](https://github.com/moesio-f/claude-code-langchain/blob/main/LICENSE)
|
|
6
|
+
|
|
7
|
+
A LangChain chat model that runs Claude through the **Claude Code CLI**, using the
|
|
8
|
+
[Claude Agent SDK](https://github.com/anthropics/claude-agent-sdk-python).
|
|
9
|
+
|
|
10
|
+
> **Not affiliated with Anthropic.** This is an independent community project.
|
|
11
|
+
> "Anthropic", "Claude", and "Claude Code" are trademarks of Anthropic PBC.
|
|
12
|
+
|
|
13
|
+
Authentication is delegated entirely to the `claude` CLI, so the model runs with whatever
|
|
14
|
+
the CLI is already configured with. For products or services that serve other users,
|
|
15
|
+
Anthropic requires API-key authentication — see the
|
|
16
|
+
[Claude Code legal and compliance page](https://code.claude.com/docs/en/legal-and-compliance).
|
|
17
|
+
|
|
18
|
+
> Built on the ideas of [ChamaruAmasara/langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code),
|
|
19
|
+
> rewritten for the current stack: the new `claude-agent-sdk` (replacing the deprecated
|
|
20
|
+
> `claude-code-sdk`) and LangChain 1.0+.
|
|
21
|
+
|
|
22
|
+
## What the rewrite changes
|
|
23
|
+
|
|
24
|
+
| Area | Original (`claude-code-sdk`) | Here (`claude-agent-sdk`) |
|
|
25
|
+
|---|---|---|
|
|
26
|
+
| SDK options | `ClaudeCodeOptions` | `ClaudeAgentOptions` |
|
|
27
|
+
| Tool calling | Tool schemas injected into the system prompt, JSON parsed out of the reply | Real Anthropic tool use: tools are served over an in-process MCP server, so `AIMessage.tool_calls` comes from actual `tool_use` blocks |
|
|
28
|
+
| Structured output | Prompted, best-effort | The CLI's `--json-schema` output format, validated against your schema |
|
|
29
|
+
| Extended thinking | A "think step by step" hint appended to the prompt | The SDK's native `thinking` config, returned as a `reasoning` content block |
|
|
30
|
+
| Effort | `--effort` via `extra_args` | First-class `effort` option, `low` → `max` |
|
|
31
|
+
| Token usage | Not available | `usage_metadata`, including cache reads/writes, plus cost in `response_metadata` |
|
|
32
|
+
| Built-in tools | Implicitly on, gated by `max_turns` | Explicit `builtin_tools`; off by default, so plain chat stays plain |
|
|
33
|
+
| LangChain | 0.3 | 1.0+ (v1 content blocks, `usage_metadata`, standard `bind_tools` / `with_structured_output`) |
|
|
34
|
+
|
|
35
|
+
## Requirements
|
|
36
|
+
|
|
37
|
+
- Python 3.10+
|
|
38
|
+
- Node.js 18+ and the Claude Code CLI: `npm install -g @anthropic-ai/claude-code && claude`
|
|
39
|
+
- The CLI authenticated with the method you intend to use (API key, Claude subscription, or
|
|
40
|
+
a supported cloud provider)
|
|
41
|
+
|
|
42
|
+
```bash
|
|
43
|
+
pip install claude-code-langchain
|
|
44
|
+
```
|
|
45
|
+
|
|
46
|
+
## Quick start
|
|
47
|
+
|
|
48
|
+
```python
|
|
49
|
+
from claude_code_langchain import ChatClaudeCode
|
|
50
|
+
|
|
51
|
+
llm = ChatClaudeCode(model="claude-opus-5")
|
|
52
|
+
print(llm.invoke("What is the capital of France?").text)
|
|
53
|
+
```
|
|
54
|
+
|
|
55
|
+
`model` accepts a model id (`claude-opus-5`, `claude-sonnet-5`, `claude-haiku-4-5`) or a CLI
|
|
56
|
+
alias (`opus`, `sonnet`, `haiku`). Leave it unset to use whatever the CLI is configured with.
|
|
57
|
+
|
|
58
|
+
### Messages, streaming, async
|
|
59
|
+
|
|
60
|
+
```python
|
|
61
|
+
from langchain_core.messages import HumanMessage, SystemMessage
|
|
62
|
+
|
|
63
|
+
llm.invoke([SystemMessage("Be concise."), HumanMessage("Explain generators.")])
|
|
64
|
+
|
|
65
|
+
for chunk in llm.stream("Count from 1 to 5"):
|
|
66
|
+
print(chunk.text, end="", flush=True)
|
|
67
|
+
|
|
68
|
+
await llm.ainvoke("Hello") # also: astream, batch, abatch
|
|
69
|
+
```
|
|
70
|
+
|
|
71
|
+
### Tool calling
|
|
72
|
+
|
|
73
|
+
```python
|
|
74
|
+
from langchain_core.tools import tool
|
|
75
|
+
|
|
76
|
+
|
|
77
|
+
@tool
|
|
78
|
+
def get_weather(city: str) -> str:
|
|
79
|
+
"""Get the weather for a city."""
|
|
80
|
+
return "12C and sunny"
|
|
81
|
+
|
|
82
|
+
|
|
83
|
+
response = llm.bind_tools([get_weather]).invoke("What is the weather in Paris?")
|
|
84
|
+
response.tool_calls
|
|
85
|
+
# [{'name': 'get_weather', 'args': {'city': 'Paris'}, 'id': 'toolu_…', 'type': 'tool_call'}]
|
|
86
|
+
```
|
|
87
|
+
|
|
88
|
+
Your tools are exposed to the CLI over an in-process MCP server, but they are never executed
|
|
89
|
+
there — the call is handed back to you, exactly like any other LangChain chat model, so it
|
|
90
|
+
works with LangGraph's `create_agent` and the usual `ToolNode` loop.
|
|
91
|
+
|
|
92
|
+
### Structured output
|
|
93
|
+
|
|
94
|
+
```python
|
|
95
|
+
from pydantic import BaseModel
|
|
96
|
+
|
|
97
|
+
|
|
98
|
+
class Person(BaseModel):
|
|
99
|
+
"""A person."""
|
|
100
|
+
|
|
101
|
+
name: str
|
|
102
|
+
age: int
|
|
103
|
+
|
|
104
|
+
|
|
105
|
+
llm.with_structured_output(Person).invoke("Bob is 42.")
|
|
106
|
+
# Person(name='Bob', age=42)
|
|
107
|
+
```
|
|
108
|
+
|
|
109
|
+
This uses the CLI's native JSON-schema output mode, so the response is constrained rather
|
|
110
|
+
than merely requested. Pass `include_raw=True` to get `{"raw", "parsed", "parsing_error"}`.
|
|
111
|
+
|
|
112
|
+
### Thinking and effort
|
|
113
|
+
|
|
114
|
+
```python
|
|
115
|
+
llm = ChatClaudeCode(
|
|
116
|
+
model="claude-opus-5",
|
|
117
|
+
thinking={"type": "adaptive", "display": "summarized"},
|
|
118
|
+
effort="high",
|
|
119
|
+
)
|
|
120
|
+
message = llm.invoke("Prove that sqrt(2) is irrational.")
|
|
121
|
+
message.content # [{'type': 'reasoning', ...}, {'type': 'text', ...}]
|
|
122
|
+
message.text # just the answer
|
|
123
|
+
```
|
|
124
|
+
|
|
125
|
+
### Agentic mode
|
|
126
|
+
|
|
127
|
+
Claude Code's own tools are **off by default**, so the model behaves like a normal chat model.
|
|
128
|
+
Turn them on when you want an agent that can touch your filesystem:
|
|
129
|
+
|
|
130
|
+
```python
|
|
131
|
+
from claude_code_langchain import READ_ONLY_TOOLS, ChatClaudeCode, ClaudeTool
|
|
132
|
+
|
|
133
|
+
reader = ChatClaudeCode(
|
|
134
|
+
builtin_tools=READ_ONLY_TOOLS, # Read, Glob, Grep
|
|
135
|
+
allowed_tools=READ_ONLY_TOOLS, # run without asking for permission
|
|
136
|
+
max_turns=10,
|
|
137
|
+
cwd="/path/to/project",
|
|
138
|
+
)
|
|
139
|
+
reader.invoke("Find every TODO comment in this project")
|
|
140
|
+
|
|
141
|
+
full = ChatClaudeCode(
|
|
142
|
+
builtin_tools="claude_code", # the complete Claude Code toolset
|
|
143
|
+
permission_mode="acceptEdits",
|
|
144
|
+
max_turns=20,
|
|
145
|
+
cwd="/path/to/project",
|
|
146
|
+
)
|
|
147
|
+
```
|
|
148
|
+
|
|
149
|
+
Presets: `READ_ONLY_TOOLS`, `WRITE_TOOLS`, `SHELL_TOOLS`, `NETWORK_TOOLS`, `ALL_TOOLS`, plus the
|
|
150
|
+
`ClaudeTool` enum for individual names.
|
|
151
|
+
|
|
152
|
+
### Sessions
|
|
153
|
+
|
|
154
|
+
Every response carries the CLI session id, which you can resume later:
|
|
155
|
+
|
|
156
|
+
```python
|
|
157
|
+
first = llm.invoke("Remember the number 7.")
|
|
158
|
+
session = first.response_metadata["session_id"]
|
|
159
|
+
|
|
160
|
+
resumed = ChatClaudeCode(session_id=session).invoke("What number did I mention?")
|
|
161
|
+
```
|
|
162
|
+
|
|
163
|
+
## Configuration
|
|
164
|
+
|
|
165
|
+
| Option | Purpose |
|
|
166
|
+
|---|---|
|
|
167
|
+
| `model`, `system_prompt` | Model id/alias; system prompt used when no `SystemMessage` is present |
|
|
168
|
+
| `thinking`, `effort` | Extended thinking config; `low` / `medium` / `high` / `xhigh` / `max` |
|
|
169
|
+
| `max_turns`, `max_budget_usd` | Agent loop limits (defaults to a single turn) |
|
|
170
|
+
| `builtin_tools`, `allowed_tools`, `disallowed_tools`, `permission_mode` | Which Claude Code tools exist, which run unattended, and how permissions are handled |
|
|
171
|
+
| `mcp_servers` | Additional MCP servers for the session |
|
|
172
|
+
| `cwd`, `add_dirs`, `env`, `cli_path`, `settings`, `setting_sources` | Where and how the CLI subprocess runs |
|
|
173
|
+
| `session_id`, `fork_session` | Resume or branch a previous session |
|
|
174
|
+
| `extra_args` | Any other CLI flag |
|
|
175
|
+
|
|
176
|
+
`setting_sources` defaults to `[]` so a run is not affected by your local Claude settings or
|
|
177
|
+
`CLAUDE.md`. Pass `None` to load user, project and local settings like the interactive CLI.
|
|
178
|
+
|
|
179
|
+
## Compared with `ChatAnthropic`
|
|
180
|
+
|
|
181
|
+
| Feature | `ChatAnthropic` | `ChatClaudeCode` |
|
|
182
|
+
|---|---|---|
|
|
183
|
+
| `invoke` / `stream` / `batch` / async | ✅ | ✅ |
|
|
184
|
+
| System messages, image input | ✅ | ✅ |
|
|
185
|
+
| Tool calling, structured output | ✅ | ✅ |
|
|
186
|
+
| Extended thinking, effort | ✅ | ✅ |
|
|
187
|
+
| Token usage and cost | ✅ | ✅ (plus USD cost per call) |
|
|
188
|
+
| Authentication | API key | Delegated to the `claude` CLI |
|
|
189
|
+
| Built-in filesystem/bash/web tools, MCP servers, subagents, skills | ❌ | ✅ |
|
|
190
|
+
| `temperature`, `top_p`, `top_k`, `stop` | ✅ | Accepted for compatibility, ignored by the CLI |
|
|
191
|
+
| Multi-turn history | Sent verbatim | Flattened into one prompt (see below) |
|
|
192
|
+
|
|
193
|
+
### Known limitations
|
|
194
|
+
|
|
195
|
+
- The CLI owns conversation state, so a message list is flattened into a single prompt with
|
|
196
|
+
labelled turns. Prior tool calls and results are rendered as text. Use `session_id` when you
|
|
197
|
+
want the CLI's own memory of a conversation instead.
|
|
198
|
+
- Sampling parameters and stop sequences are not exposed by the CLI.
|
|
199
|
+
- Streaming yields text only; thinking arrives on the final non-streamed message.
|
|
200
|
+
- Each call spawns a `claude` subprocess, so latency is higher than a direct API call.
|
|
201
|
+
|
|
202
|
+
## Development
|
|
203
|
+
|
|
204
|
+
```bash
|
|
205
|
+
uv sync --group dev
|
|
206
|
+
uv run pytest # unit tests, no CLI needed
|
|
207
|
+
uv run ruff check . && uv run mypy src
|
|
208
|
+
CLAUDE_CODE_LANGCHAIN_INTEGRATION=1 uv run pytest tests/integration # runs the CLI
|
|
209
|
+
```
|
|
210
|
+
|
|
211
|
+
## Related projects
|
|
212
|
+
|
|
213
|
+
Other LangChain adapters for Claude Code:
|
|
214
|
+
|
|
215
|
+
| Project | Notes |
|
|
216
|
+
|---|---|
|
|
217
|
+
| [ChamaruAmasara/langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code) | The original project this one was inspired by; on PyPI as `langchain-claude-code-cli` |
|
|
218
|
+
| [thehumanworks/langchain-claude-code](https://github.com/thehumanworks/langchain-claude-code) | Wrapper for the Claude Agent SDK; on PyPI as `langchain-claude-code` |
|
|
219
|
+
| [brhodes/langchain-claude-agent-sdk](https://codeberg.org/brhodes/langchain-claude-agent-sdk) | `langchain-claude-agent-sdk` on PyPI (hosted on Codeberg rather than GitHub) |
|
|
220
|
+
| [langchain-ai/langchain#40131](https://github.com/langchain-ai/langchain/issues/40131) | Proposal to add an Agent SDK-backed model to `langchain-anthropic` |
|
|
221
|
+
|
|
222
|
+
Each takes a different approach to authentication and conversation state. Some accept OAuth
|
|
223
|
+
tokens directly; review the auth method and Anthropic's terms before using one in production.
|
|
224
|
+
|
|
225
|
+
## Credits
|
|
226
|
+
|
|
227
|
+
Original project: [langchain-claude-code](https://github.com/ChamaruAmasara/langchain-claude-code)
|
|
228
|
+
by Chamaru Amasara. This package is an independent rewrite on the Claude Agent SDK and
|
|
229
|
+
LangChain 1.0+.
|
|
@@ -0,0 +1,86 @@
|
|
|
1
|
+
[project]
|
|
2
|
+
name = "claude-code-langchain"
|
|
3
|
+
version = "0.1.0"
|
|
4
|
+
description = "LangChain chat model for Claude Code, built on the Claude Agent SDK"
|
|
5
|
+
license = "MIT"
|
|
6
|
+
license-files = ["LICENSE"]
|
|
7
|
+
readme = "README.md"
|
|
8
|
+
requires-python = ">=3.10"
|
|
9
|
+
keywords = [
|
|
10
|
+
"langchain",
|
|
11
|
+
"claude",
|
|
12
|
+
"anthropic",
|
|
13
|
+
"claude-code",
|
|
14
|
+
"claude-agent-sdk",
|
|
15
|
+
]
|
|
16
|
+
classifiers = [
|
|
17
|
+
"Development Status :: 3 - Alpha",
|
|
18
|
+
"Intended Audience :: Developers",
|
|
19
|
+
"Programming Language :: Python :: 3",
|
|
20
|
+
"Programming Language :: Python :: 3.10",
|
|
21
|
+
"Programming Language :: Python :: 3.11",
|
|
22
|
+
"Programming Language :: Python :: 3.12",
|
|
23
|
+
"Programming Language :: Python :: 3.13",
|
|
24
|
+
"Topic :: Software Development :: Libraries",
|
|
25
|
+
]
|
|
26
|
+
dependencies = [
|
|
27
|
+
"claude-agent-sdk>=0.2.0",
|
|
28
|
+
"langchain-core>=1.0.0,<2.0.0",
|
|
29
|
+
]
|
|
30
|
+
|
|
31
|
+
[[project.authors]]
|
|
32
|
+
name = "Moésio Filho"
|
|
33
|
+
email = "moesio.f@gmail.com"
|
|
34
|
+
|
|
35
|
+
[project.urls]
|
|
36
|
+
Homepage = "https://github.com/moesio-f/claude-code-langchain"
|
|
37
|
+
Repository = "https://github.com/moesio-f/claude-code-langchain"
|
|
38
|
+
Issues = "https://github.com/moesio-f/claude-code-langchain/issues"
|
|
39
|
+
|
|
40
|
+
[build-system]
|
|
41
|
+
requires = ["uv_build>=0.11.30,<0.12.0"]
|
|
42
|
+
build-backend = "uv_build"
|
|
43
|
+
|
|
44
|
+
[dependency-groups]
|
|
45
|
+
dev = [
|
|
46
|
+
"mypy>=1.18",
|
|
47
|
+
"pytest>=8.4",
|
|
48
|
+
"pytest-asyncio>=1.0",
|
|
49
|
+
"ruff>=0.13",
|
|
50
|
+
]
|
|
51
|
+
|
|
52
|
+
[tool.pytest.ini_options]
|
|
53
|
+
asyncio_mode = "auto"
|
|
54
|
+
testpaths = ["tests"]
|
|
55
|
+
|
|
56
|
+
[tool.ruff.lint]
|
|
57
|
+
select = [
|
|
58
|
+
"E",
|
|
59
|
+
"F",
|
|
60
|
+
"I",
|
|
61
|
+
"UP",
|
|
62
|
+
"B",
|
|
63
|
+
"SIM",
|
|
64
|
+
"D",
|
|
65
|
+
]
|
|
66
|
+
ignore = [
|
|
67
|
+
"D100",
|
|
68
|
+
"D107",
|
|
69
|
+
"D203",
|
|
70
|
+
"D213",
|
|
71
|
+
]
|
|
72
|
+
|
|
73
|
+
[tool.ruff.lint.pydocstyle]
|
|
74
|
+
convention = "google"
|
|
75
|
+
|
|
76
|
+
[tool.ruff.lint.per-file-ignores]
|
|
77
|
+
"tests/**" = ["D"]
|
|
78
|
+
|
|
79
|
+
[tool.mypy]
|
|
80
|
+
plugins = ["pydantic.mypy"]
|
|
81
|
+
disallow_untyped_defs = true
|
|
82
|
+
warn_return_any = false
|
|
83
|
+
|
|
84
|
+
[[tool.mypy.overrides]]
|
|
85
|
+
module = "tests.*"
|
|
86
|
+
disallow_untyped_defs = false
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
[project]
|
|
2
|
+
name = "claude-code-langchain"
|
|
3
|
+
version = "0.1.0"
|
|
4
|
+
description = "LangChain chat model for Claude Code, built on the Claude Agent SDK"
|
|
5
|
+
license = "MIT"
|
|
6
|
+
license-files = ["LICENSE"]
|
|
7
|
+
readme = "README.md"
|
|
8
|
+
authors = [
|
|
9
|
+
{ name = "Moésio Filho", email = "moesio.f@gmail.com" }
|
|
10
|
+
]
|
|
11
|
+
requires-python = ">=3.10"
|
|
12
|
+
keywords = ["langchain", "claude", "anthropic", "claude-code", "claude-agent-sdk"]
|
|
13
|
+
classifiers = [
|
|
14
|
+
"Development Status :: 3 - Alpha",
|
|
15
|
+
"Intended Audience :: Developers",
|
|
16
|
+
"Programming Language :: Python :: 3",
|
|
17
|
+
"Programming Language :: Python :: 3.10",
|
|
18
|
+
"Programming Language :: Python :: 3.11",
|
|
19
|
+
"Programming Language :: Python :: 3.12",
|
|
20
|
+
"Programming Language :: Python :: 3.13",
|
|
21
|
+
"Topic :: Software Development :: Libraries",
|
|
22
|
+
]
|
|
23
|
+
dependencies = [
|
|
24
|
+
"claude-agent-sdk>=0.2.0",
|
|
25
|
+
"langchain-core>=1.0.0,<2.0.0",
|
|
26
|
+
]
|
|
27
|
+
|
|
28
|
+
[project.urls]
|
|
29
|
+
Homepage = "https://github.com/moesio-f/claude-code-langchain"
|
|
30
|
+
Repository = "https://github.com/moesio-f/claude-code-langchain"
|
|
31
|
+
Issues = "https://github.com/moesio-f/claude-code-langchain/issues"
|
|
32
|
+
|
|
33
|
+
[build-system]
|
|
34
|
+
requires = ["uv_build>=0.11.30,<0.12.0"]
|
|
35
|
+
build-backend = "uv_build"
|
|
36
|
+
|
|
37
|
+
[dependency-groups]
|
|
38
|
+
dev = [
|
|
39
|
+
"mypy>=1.18",
|
|
40
|
+
"pytest>=8.4",
|
|
41
|
+
"pytest-asyncio>=1.0",
|
|
42
|
+
"ruff>=0.13",
|
|
43
|
+
]
|
|
44
|
+
|
|
45
|
+
[tool.pytest.ini_options]
|
|
46
|
+
asyncio_mode = "auto"
|
|
47
|
+
testpaths = ["tests"]
|
|
48
|
+
|
|
49
|
+
[tool.ruff.lint]
|
|
50
|
+
select = ["E", "F", "I", "UP", "B", "SIM", "D"]
|
|
51
|
+
ignore = ["D100", "D107", "D203", "D213"]
|
|
52
|
+
pydocstyle.convention = "google"
|
|
53
|
+
|
|
54
|
+
[tool.ruff.lint.per-file-ignores]
|
|
55
|
+
"tests/**" = ["D"]
|
|
56
|
+
|
|
57
|
+
[tool.mypy]
|
|
58
|
+
plugins = ["pydantic.mypy"]
|
|
59
|
+
disallow_untyped_defs = true
|
|
60
|
+
warn_return_any = false
|
|
61
|
+
|
|
62
|
+
[[tool.mypy.overrides]]
|
|
63
|
+
module = "tests.*"
|
|
64
|
+
disallow_untyped_defs = false
|