toolfunnel 0.4.0 → 0.5.0
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 +167 -342
- package/docs/KNOWN_BUGS.md +24 -216
- package/docs/MANUAL.pdf +0 -0
- package/package.json +1 -1
- package/src/mcp/http-transport.js +31 -0
- package/src/mcp/server.js +98 -15
- package/src/tools/registry.js +4 -2
- package/docs/admin-ui.md +0 -404
- package/docs/configuration.md +0 -402
- package/docs/hooks-and-gating.md +0 -428
- package/docs/logging.md +0 -223
- package/docs/management.md +0 -429
- package/docs/oauth.md +0 -103
- package/docs/releasing.md +0 -31
- package/docs/setup.md +0 -352
- package/docs/tools-and-upstreams.md +0 -469
package/README.md
CHANGED
|
@@ -1,342 +1,167 @@
|
|
|
1
|
-
# ToolFunnel
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
|
|
5
|
-
> A zero-dependency
|
|
6
|
-
|
|
7
|
-
**Zero code, zero dependencies: your own MCP server in 60 seconds** — from three ordinary scripts to a named, packaged, policy-gated MCP server (everything on screen is real output; see [demo/](demo/)):
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
### The
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
The
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
|
79
|
-
|
|
80
|
-
|
|
|
81
|
-
|
|
|
82
|
-
|
|
|
83
|
-
|
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
-
|
|
112
|
-
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
159
|
-
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
|
|
167
|
-
|
|
168
|
-
- **Tame context bloat.** Several rich MCP servers connected at once drown the agent's context in tool schemas it won't use this turn. Behind ToolFunnel the agent sees short briefs and loads a tool's full schema only when it actually reaches for it.
|
|
169
|
-
- **One toolbox, many workflows.** Keep all your multi-use tools and common MCP servers in one place and pick which to surface per workflow - instead of re-wiring a different setup for every client.
|
|
170
|
-
- **Govern what an agent may run - on any client.** Put a fail-closed PreToolUse policy *in the gateway* so it travels with your tools, even to clients that have no hook system of their own.
|
|
171
|
-
- **Turn a script into an MCP tool without building an MCP.** Have a useful CLI, a bash one-liner, or a script in any language? Declare it in a JSON entry and it's a gated MCP tool - no protocol code, no SDK (`toolfunnel_howto({ topic: "create-tool" })` shows the shapes).
|
|
172
|
-
- **Add a tool mid-session.** Need a tool while you're working? Drop it in and it's live on the next turn - no restart.
|
|
173
|
-
- **Survive an upstream crash.** If an attached MCP server dies mid-session, ToolFunnel notices, reconnects it in the background with backoff, and re-advertises its tools - no restart, no dropped session. Most reconnect logic elsewhere is Docker-scoped; this isn't.
|
|
174
|
-
|
|
175
|
-
## Zero dependencies, on purpose
|
|
176
|
-
|
|
177
|
-
ToolFunnel has **no runtime npm dependencies** (`"dependencies": {}`) and does **not** use an MCP SDK - the JSON-RPC 2.0 wire protocol is hand-rolled on top of Node built-ins (`node:http`, `node:child_process`, `node:fs`, …). Requires Node **>= 18**. `npm install toolfunnel` pulls **zero** packages.
|
|
178
|
-
|
|
179
|
-
> **\* The one asterisk: OAuth is opt-in.** Enabling OAuth 2.1 (off by default) adds exactly **one** dependency - [`jose`](https://www.npmjs.com/package/jose), which is itself **zero-dependency**, audited, and the same library the official MCP SDK uses. It is **not** a runtime dependency of the core: it appears only as a `devDependency` (for the OAuth test suite) and is **installed on demand** for users who turn auth on - `toolfunnel install-oauth` or the Install button in the UI's Auth tab. So the default footprint is genuinely zero, and the most security-sensitive code in the project - token validation - is delegated to an audited library rather than hand-rolled. See [Authentication](#authentication---optional-oauth-21).
|
|
180
|
-
|
|
181
|
-
This is a deliberate **security** decision, not minimalism for its own sake. ToolFunnel sits in a privileged position - it **gates tool execution** - which is exactly where you don't want an unaudited dependency tree. And `dependencies: {}` in npm is a stronger claim than a "single binary" elsewhere: a Go gateway that ships as one binary still *statically links* its dependency tree (mcpproxy-go's, for example, runs to 40+ direct deps - goja, esbuild, gRPC, an observability stack). That's dependency-**bundled**, not dependency-free. In a year when npm itself saw large packages compromised (Axios, the 140-package Mastra incident), a tool-execution gate with **nothing transitive to audit** is a defensible engineering stance - you audit *your* code, not forty supply chains. (It is not a silver bullet: a zero-dep posture shrinks the audit surface, it doesn't remove the burden of getting the hand-rolled wire + gate correct - which I have tried to do.)
|
|
182
|
-
|
|
183
|
-
The transport reads **both** LSP-style `Content-Length:` framing **and** newline-delimited JSON, and writes newline-delimited JSON, so it interoperates with simple and framed clients alike.
|
|
184
|
-
|
|
185
|
-
## Authentication - optional OAuth 2.1
|
|
186
|
-
|
|
187
|
-
By default the gateway is **loopback-only and unauthenticated** - the right posture for a single operator on `localhost`. For a networked or multi-user deployment, ToolFunnel can act as an **OAuth 2.1 resource server**: it validates the bearer token on every request before any tool runs.
|
|
188
|
-
|
|
189
|
-
- **Opt-in, one dependency.** Off until you enable it. Enabling installs `jose` on demand (`toolfunnel install-oauth` or the UI Auth tab) and the core stays zero-runtime-dependency for everyone else.
|
|
190
|
-
- **Delegated, not hand-rolled.** Token validation goes through `jose` with a **pinned algorithm allowlist** (which defeats the RS256↔HS256 confusion attack and `alg:none`), enforced **issuer**, and an enforced **audience** bound to this gateway's resource URI - the RFC 8707 confused-deputy defence. JWKS fetch, caching, and key rotation are the library's job.
|
|
191
|
-
- **Discovery.** With auth on, the gateway serves the RFC 9728 Protected Resource Metadata document at `GET /.well-known/oauth-protected-resource` (unauthenticated), and a `401` carries a `WWW-Authenticate: Bearer resource_metadata="…"` hint.
|
|
192
|
-
- **Safe exposure.** Auth is what unlocks a non-loopback bind: the HTTP host **refuses to bind off-localhost unless OAuth is enabled**, and refuses to start if auth is on but the dependency is missing or the config is incoherent (fail-fast).
|
|
193
|
-
- **Configure** in the UI Auth tab or `auth/auth.config.json` (`enabled`, `issuer`, `audience`, optional `jwksUri` - else derived from the issuer via OIDC discovery - `algorithms`, `requiredScopes`, `clockToleranceSec`). Requires Node **>= 18** (`jose` v5, the CommonJS build).
|
|
194
|
-
|
|
195
|
-
The OAuth **client** leg (Dynamic Client Registration, authorization-server-metadata discovery, step-up auth) and the 2026 spec hardening (RFC 9207 issuer validation, credential binding) are **planned, not yet shipped** - the resource-server slice is the shipped, tested MVP.
|
|
196
|
-
|
|
197
|
-
## Token efficiency
|
|
198
|
-
|
|
199
|
-
The mechanism is the lean register: a connected upstream normally injects every tool's full JSON schema into context on every turn. ToolFunnel replaces that with short briefs (`name`, `summary`, `category`) and serves the full instructions for a single tool only when the agent asks via `toolfunnel_tool_instructions`. The more tools you connect, the larger the saving.
|
|
200
|
-
|
|
201
|
-
## Quickstart
|
|
202
|
-
|
|
203
|
-
```bash
|
|
204
|
-
# 1. Get the code
|
|
205
|
-
git clone <your-fork-or-clone-url> toolfunnel
|
|
206
|
-
cd toolfunnel
|
|
207
|
-
|
|
208
|
-
# 2. "Install" - zero RUNTIME deps; a clone's npm install pulls only the dev/test
|
|
209
|
-
# tooling (jose, for the OAuth test suite). Installed as a dependency it pulls nothing.
|
|
210
|
-
npm install # runtime dependencies: none
|
|
211
|
-
|
|
212
|
-
# 3a. Run as a stdio MCP server (the default - what most clients spawn)
|
|
213
|
-
node bin/toolfunnel.js
|
|
214
|
-
|
|
215
|
-
# 3b. Or run as an HTTP/SSE host on 127.0.0.1:9998 (long-lived; many clients can connect)
|
|
216
|
-
node bin/toolfunnel.js --http
|
|
217
|
-
node bin/toolfunnel.js --http --port 0 # 0 = OS-assigned port
|
|
218
|
-
node bin/toolfunnel.js --http --host 127.0.0.1 --port 9998
|
|
219
|
-
|
|
220
|
-
# 3c. Or open the OPTIONAL config web UI on 127.0.0.1:9777 (loopback only)
|
|
221
|
-
node bin/toolfunnel.js --ui
|
|
222
|
-
node bin/toolfunnel.js --ui --port 0 # 0 = OS-assigned port
|
|
223
|
-
|
|
224
|
-
# Help
|
|
225
|
-
node bin/toolfunnel.js --help
|
|
226
|
-
|
|
227
|
-
# Smoke-test the wiring (build → initialize → tools/list → call a meta-tool)
|
|
228
|
-
node test/smoke.js
|
|
229
|
-
```
|
|
230
|
-
|
|
231
|
-
The HTTP host binds loopback only and exposes a single Streamable-HTTP endpoint at `POST/GET /mcp` (plus a deprecated `GET /mcp/sse` alias and a `GET /health` JSON snapshot). The npm scripts mirror the above: `npm start` (stdio) and `npm run http`.
|
|
232
|
-
|
|
233
|
-
### Registering with an MCP client
|
|
234
|
-
|
|
235
|
-
Point your client at ToolFunnel via `.mcp.json`. For a **stdio** server the client spawns the process:
|
|
236
|
-
|
|
237
|
-
```json
|
|
238
|
-
{ "mcpServers": { "toolfunnel": { "command": "node", "args": ["/path/to/toolfunnel/bin/toolfunnel.js"] } } }
|
|
239
|
-
```
|
|
240
|
-
|
|
241
|
-
For the **HTTP** host (start it first with `node bin/toolfunnel.js --http`), point the client at the `/mcp` endpoint:
|
|
242
|
-
|
|
243
|
-
```json
|
|
244
|
-
{ "mcpServers": { "toolfunnel": { "type": "http", "url": "http://127.0.0.1:9998/mcp" } } }
|
|
245
|
-
```
|
|
246
|
-
|
|
247
|
-
Once connected, call `toolfunnel_list_tools` to see the briefs, `toolfunnel_tool_instructions` with a tool `name` to get its full usage, and `toolfunnel_howto` to learn how to extend the gateway.
|
|
248
|
-
|
|
249
|
-
**Or skip the config entirely - just ask your AI.** Point any MCP-aware client at ToolFunnel and say *"attach this tool server, expose these tools, and block anything destructive."* The AI reads ToolFunnel's own built-in instructions and wires it all up for you - upstreams, exposed tools, and a PreToolUse safety gate - in plain language, no JSON editing. Prefer to click? The web UI below does the same. **No coding experience required :)**
|
|
250
|
-
|
|
251
|
-
## The config web UI
|
|
252
|
-
|
|
253
|
-
An **optional** local admin console lets a human view, search, and configure the gateway without hand-editing JSON:
|
|
254
|
-
|
|
255
|
-
```bash
|
|
256
|
-
node bin/toolfunnel.js --ui # http://127.0.0.1:9777
|
|
257
|
-
node bin/toolfunnel.js --ui --port 0 # OS-assigned port
|
|
258
|
-
```
|
|
259
|
-
|
|
260
|
-
It binds **loopback only**, rejects a non-loopback `Host` header (DNS-rebind guard), and has zero front-end dependencies (vanilla HTML/CSS/JS, dark theme, offline - no CDN, no framework). Five tabs:
|
|
261
|
-
|
|
262
|
-
- **Tools** - every register tool with live search and a **show-hidden** filter; a top-level **surface panel** (the four meta-tools' every-turn toggles + a promotion count and footgun warnings); add a tool (script body, shell, or invoke-less reference); toggle **Enabled** (lean), **Hot** (every-turn promotion), and **Hidden** (declutter this manager list); flip per-tool **execution mode** (reference ↔ gateway); toggle per-tool **Pre/Post hook** gates; remove.
|
|
263
|
-
- **MCPs** - upstream MCP servers and their curated-direct expose selections; add an upstream with exposed tools; enable / disable / remove; **Discover** - a live connect-and-list of an upstream's tools, each with its own **lean** (enabled) and **Hot** toggle keyed by the surfaced name.
|
|
264
|
-
- **Hooks** - the hook manifest with live enabled state; add a lifecycle hook (optionally authoring its script body); enable / disable / remove.
|
|
265
|
-
- **Logs** - the on/off switch for the [audit log](#activity--audit-log) and a newest-first view of recent records.
|
|
266
|
-
- **Auth** - configure the optional OAuth 2.1 resource-server validation (off by default) and install its single dependency on demand.
|
|
267
|
-
|
|
268
|
-
Every UI write goes through the **same stores** the MCP server reads, so an edit is byte-identical to a hand edit (and a CLI edit) and is visible to the running server with no restart.
|
|
269
|
-
|
|
270
|
-
## Activity & audit log
|
|
271
|
-
|
|
272
|
-
ToolFunnel can keep a **JSONL** activity/audit log - **off by default** (a fresh checkout writes nothing and creates no log file). When enabled, it records tool runs and **every gate allow/deny decision**. Toggle it three ways, all equivalent:
|
|
273
|
-
|
|
274
|
-
- the `tf_log` management function - `{ action: "enable" | "disable" | "status" }`;
|
|
275
|
-
- the web UI **Logs** tab switch;
|
|
276
|
-
- editing `logs/log.config.json` directly (`{ "enabled": bool, "path": "logs/toolfunnel.log.jsonl" }`).
|
|
277
|
-
|
|
278
|
-
Config is read **fresh per event**, so a toggle takes effect immediately. Logging is wrapped so it can **never** throw into a tool call or the gate, and `tf_log` can itself be disabled to revoke logging control. (The `logs/` directory is git-ignored.)
|
|
279
|
-
|
|
280
|
-
## Project layout
|
|
281
|
-
|
|
282
|
-
ToolFunnel keeps a deliberate split between **the engine** (`src/`, which you never edit) and **what you manage** (`tools/`, `mcp/`, `hooks/` - plain config and scripts at the top level):
|
|
283
|
-
|
|
284
|
-
```
|
|
285
|
-
toolfunnel/
|
|
286
|
-
├─ bin/toolfunnel.js # entry point - stdio by default, --http for the host, --ui for the console
|
|
287
|
-
├─ src/ # ToolFunnel's own engine (you never edit this)
|
|
288
|
-
│ ├─ mcp/ # server, protocol, transports, mcp-client, aggregator, gated-run
|
|
289
|
-
│ ├─ core/ # hook-engine, hook-loader, hook-runner, matcher, events, logger
|
|
290
|
-
│ ├─ tools/ # the tool SYSTEM (registry, tool-state, drift)
|
|
291
|
-
│ ├─ ui/ # the optional config web UI (loopback server + vanilla assets)
|
|
292
|
-
│ └─ extend/ · packages/
|
|
293
|
-
├─ tools/ # YOUR tools
|
|
294
|
-
│ ├─ tools.register.json # first-party tool register (7 demos + 8 management functions)
|
|
295
|
-
│ ├─ tools.state.json # enable/disable overlay (default-ON)
|
|
296
|
-
│ └─ scripts/ # the tool scripts
|
|
297
|
-
├─ mcp/ # YOUR upstreams
|
|
298
|
-
│ ├─ expose.json # upstream servers + curated tools to expose (empty default)
|
|
299
|
-
│ └─ expose.example.json # annotated sample
|
|
300
|
-
├─ hooks/ # YOUR policy gate
|
|
301
|
-
│ ├─ hooks.manifest.json # the gate (empty = allow-all)
|
|
302
|
-
│ ├─ hooks.state.json # live enable/disable overlay (optional; wins over the manifest seed)
|
|
303
|
-
│ └─ scripts/ # hook command scripts
|
|
304
|
-
└─ logs/ # audit log (git-ignored; created only when logging is enabled)
|
|
305
|
-
├─ log.config.json # { enabled, path } - default OFF
|
|
306
|
-
└─ toolfunnel.log.jsonl # the JSONL records (default path)
|
|
307
|
-
```
|
|
308
|
-
|
|
309
|
-
Rule of thumb: `src/` is the gateway machinery; `tools/`, `mcp/`, and `hooks/` are where your configuration lives. By default these resolve relative to the repo root — or point them anywhere with a **config home** (`--config-dir` / `TOOLFUNNEL_HOME`): an empty home is seeded from the shipped defaults on first use and never overwritten after, so an `npm update` can never eat your tools. See [docs/packaging.md](docs/packaging.md).
|
|
310
|
-
|
|
311
|
-
## Configuration
|
|
312
|
-
|
|
313
|
-
All config is plain JSON under the repo - edit it by hand, through the [config web UI](#the-config-web-ui), or via the [management functions](#management-functions):
|
|
314
|
-
|
|
315
|
-
| File | Purpose |
|
|
316
|
-
|-----------------------------------|-------------------------------------------------------------------------|
|
|
317
|
-
| `tools/tools.register.json` | The first-party tool register (`{ version, description, tools: [...] }`) - 7 demos + 9 management functions. Entries may carry an `inputSchema`, advertised verbatim when the tool is promoted hot |
|
|
318
|
-
| `tools/tools.state.json` | The visibility-matrix overlay, keyed by surfaced name → `{ enabled?, hidden?, hot? }`. Default-ON for `enabled` (empty `{}` enables everything), default-OFF for `hot`/`hidden` (meta-tools default `hot`-ON). Read fresh per call - no restart needed |
|
|
319
|
-
| `hooks/hooks.manifest.json` | The policy gate (`{ version, hooks: [] }`). An empty `hooks` array = allow-all |
|
|
320
|
-
| `hooks/hooks.state.json` | Optional live enable/disable overlay for hooks; an entry here wins over the manifest's seed `enabled` |
|
|
321
|
-
| `mcp/expose.json` | Upstream MCP servers + the curated set of their tools to expose (`{ version, upstreams: [], expose: [] }`). **Empty by default**, so the gateway connects to nothing |
|
|
322
|
-
| `mcp/expose.example.json` | An annotated sample showing how to wire an upstream and expose two of its tools |
|
|
323
|
-
| `logs/log.config.json` | The audit-log toggle (`{ enabled, path }`). **Default OFF**; created only when logging is enabled |
|
|
324
|
-
| `toolfunnel.json` | Optional identity + defaults: `{ serverName, serverVersion, httpPort, uiPort, requires }`. Absent = the gateway is plain `toolfunnel`. This is how a packaged MCP introduces itself as ITSELF |
|
|
325
|
-
|
|
326
|
-
## Packaging: ship your own MCP
|
|
327
|
-
|
|
328
|
-
Everything you build here — tools, curation, upstream selections, policy hooks, identity — lives
|
|
329
|
-
in one folder (the config home), and one management call packages it:
|
|
330
|
-
|
|
331
|
-
- `tf_pack { format: "home" }` → a portable setup: zip it, git it, or run it with `--config-dir`.
|
|
332
|
-
- `tf_pack { format: "npm", name: "my-mcp" }` → a **publishable npm package** that depends on
|
|
333
|
-
toolfunnel (never forks it), bundles your setup, and gives your users `npx my-mcp` — your
|
|
334
|
-
server, your name in the handshake, your tools and schemas, **your gate enforced on their
|
|
335
|
-
machine regardless of client**.
|
|
336
|
-
|
|
337
|
-
Declare runtime needs (`python`, `git`, …) in `toolfunnel.json` `requires` and the gateway warns
|
|
338
|
-
at startup about anything missing. Full story: [docs/packaging.md](docs/packaging.md).
|
|
339
|
-
|
|
340
|
-
## License
|
|
341
|
-
|
|
342
|
-
[MIT](LICENSE)
|
|
1
|
+
# ToolFunnel
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
|
|
5
|
+
> A zero-dependency MCP and Local tool gateway: host your own tools in any language, forward and curate tools from other MCP servers, expose them leanly to cut agent token cost, and gate every call through your own policy hooks before it runs + create your own MCP servers easily with zero code!
|
|
6
|
+
|
|
7
|
+
**Zero code, zero dependencies: your own MCP server in 60 seconds** — from three ordinary scripts to a named, packaged, policy-gated MCP server (everything on screen is real output; see [demo/](demo/)):
|
|
8
|
+
|
|
9
|
+
|
|
10
|
+
|
|
11
|
+
**📖 Full documentation: [the ToolFunnel User Manual](docs/MANUAL.pdf)** - installation, the web UI, writing tools, attaching MCP servers, the policy gate, packaging, and troubleshooting, all illustrated with screenshots and worked examples. This README is the overview; the manual is the reference.
|
|
12
|
+
|
|
13
|
+
## The problem
|
|
14
|
+
|
|
15
|
+
The Model Context Protocol (MCP) lets an AI agent call tools from many servers. But every connected server dumps **all** of its tool schemas into the model's context on every turn. Connect a handful of rich MCP servers and you've spent thousands of tokens describing tools the agent won't use this turn - slower, costlier, noisier.
|
|
16
|
+
|
|
17
|
+
Similarly, working with AI I found myself generating many many tools, some of which I wanted to use with different AI workflows, some I didnt, and I didnt want to keep setting them up, I also didnt wish to keep many different setups for different workflows. Therefore ToolFunnel was born so that all of my multi-use tools and common MCP servers can be in one place, with any workflow, and I can easily select which I want to use with whatever workflow. You can also change or add tools during a session with simple toggles in the UI. I also wanted to wire up and test an MCP *live*, in the running session, without resetting the CLI or restarting anything - and have its tools show up in the tools list straight away; ToolFunnel does exactly that.
|
|
18
|
+
|
|
19
|
+
Often also, there's also no consistent way to **govern** what an agent may run: hooks and policies live in the *host* (a specific CLI or otherwise), so they don't travel when you switch clients - with ToolFunnel, this is easy because your hooks travel with the tools - it can become your one swiss-army knife for many different workflows.
|
|
20
|
+
|
|
21
|
+
Finally, I didnt want to audit huge numbers of dependencies - a personal choice, yes - so I wanted something that could be audited quickly and easily.
|
|
22
|
+
|
|
23
|
+
At the time of writing, I have 14 MCP tools and 93 local tools all accessed through toolfunnel, gated as required.
|
|
24
|
+
|
|
25
|
+
## What ToolFunnel does
|
|
26
|
+
|
|
27
|
+
ToolFunnel is one small MCP server that sits between your agent and everything else:
|
|
28
|
+
|
|
29
|
+
1. **Hosts your own tools** - define first-party tools in a JSON register and serve them directly. Seven demo tools ship in the box.
|
|
30
|
+
2. **Forwards other MCP servers - leanly.** Attach an upstream MCP and its tools appear in the same lean register as your own, runnable through the gate. Curate which appear, promote a chosen few to top-level "every-turn" tools, or leave them lean by default.
|
|
31
|
+
3. **Lean register** - the agent sees short tool *briefs*; the full instructions for a tool are fetched on demand, so context stays small. *(This is the token saver.)*
|
|
32
|
+
4. **Server-side policy gate** - every server-side execution path fires your PreToolUse / PostToolUse hooks inside the gateway, so your policy works on any client, not just hosts that support hooks. The gate travels with the gateway, and fails closed.
|
|
33
|
+
5. **Configure by file, UI, or in-band** - plain JSON files, an optional loopback web UI, or nine in-band management functions (`tf_*`) all add / curate / toggle tools, upstreams, and hooks - live, no restart - and package the whole setup for deployment.
|
|
34
|
+
6. **Audit when you want it** - a toggleable JSONL log (default off) records tool runs, every gate allow/deny decision, and every upstream connect / disconnect / reconnect.
|
|
35
|
+
7. **Live & self-healing** - attach/curate/toggle on a running gateway with no restart; if an attached MCP's process dies, the gateway detects it and reconnects in the background with backoff.
|
|
36
|
+
8. **Build your own MCP server, no code, no SDK** - assemble scripts and curated upstream tools, switch the meta-tools off, and ToolFunnel *is* your MCP server; `tf_pack` ships it as a `npx`-installable npm package.
|
|
37
|
+
|
|
38
|
+
## Build your own MCP server - no code, no SDK
|
|
39
|
+
|
|
40
|
+
ToolFunnel is a no-code MCP builder. Point it at a few scripts (any language), curate a set of tools from other MCP servers, or mix both, then switch its own four meta-tools off - and what your agent connects to is a plain, top-level MCP server presenting exactly your chosen tools. No decorators, no framework, no `@tool` boilerplate, no Python. One `tf_pack` call turns it into a publishable npm package your users install with `npx your-mcp`.
|
|
41
|
+
|
|
42
|
+
That puts ToolFunnel head to head with a framework like FastMCP for the "I just want to stand up an MCP server" job, from the opposite direction:
|
|
43
|
+
|
|
44
|
+
- **FastMCP:** write Python, decorate your functions, ship a framework dependency.
|
|
45
|
+
- **ToolFunnel:** declare your tools in JSON (or let the AI author them for you), and you have a gated MCP server - assembled from scripts *and* other people's MCP servers - with zero runtime dependencies and a policy gate built in.
|
|
46
|
+
|
|
47
|
+
If you can write a script, you can ship an MCP. The full walkthrough is in the [manual](docs/MANUAL.pdf); the short version is `toolfunnel_howto({ topic: "create-tool" })` and `tf_pack`.
|
|
48
|
+
|
|
49
|
+
### The meta-tool surface
|
|
50
|
+
|
|
51
|
+
The model never sees your long tail of tools directly. It sees four fixed **meta-tools** - `toolfunnel_list_tools`, `toolfunnel_tool_instructions`, `toolfunnel_howto`, `toolfunnel_run_tool` - and reaches the real tools through them: list the briefs, read one tool's instructions on demand, then run it through the gate. Any curated upstream tools you expose are advertised alongside them and run through the same gate.
|
|
52
|
+
|
|
53
|
+
### The visibility matrix
|
|
54
|
+
|
|
55
|
+
Every tool - your own, a forwarded upstream tool, and the four meta-tools - has three independent visibility dials: **enabled** (lean-visible), **hot** (promoted to the every-turn `tools/list` surface, directly callable), and **hidden** (manager-list declutter). Toggles are live, no restart.
|
|
56
|
+
|
|
57
|
+
This is what lets ToolFunnel be a lean register and a conventional MCP at the same time: keep the long tail lean, promote the few tools you call constantly, or promote your chosen set and switch the meta-tools off to present a plain top-level MCP server (the no-code-MCP posture above). Two footguns the UI warns about: hiding the meta-tools removes the agent's ability to discover tools by name, and promoting many tools reintroduces the context bloat the lean register exists to avoid.
|
|
58
|
+
|
|
59
|
+
### The gate
|
|
60
|
+
|
|
61
|
+
Every server-side run path - your tools, the management functions, and every forwarded upstream call - funnels through one gate:
|
|
62
|
+
|
|
63
|
+
```
|
|
64
|
+
gatedRun({ engine, ctx, toolName, args, execute })
|
|
65
|
+
→ fire PreToolUse (may BLOCK; fails CLOSED if the engine errors)
|
|
66
|
+
→ execute() (ONLY if allowed)
|
|
67
|
+
→ fire PostToolUse (advisory; cannot un-run the tool)
|
|
68
|
+
```
|
|
69
|
+
|
|
70
|
+
Hooks speak the Claude-Code hook protocol (event JSON on stdin, exit `2` to block), so an existing hook travels here unchanged. The load-bearing invariant, proven by test: a PreToolUse deny means `execute()` is never called. Writing hooks, matchers, and worked examples: [manual](docs/MANUAL.pdf).
|
|
71
|
+
|
|
72
|
+
## Why it's different
|
|
73
|
+
|
|
74
|
+
Hosting tools and proxying other MCP servers is a crowded space - but for my own use case, which is why I rolled my own solution, I wanted something different.
|
|
75
|
+
|
|
76
|
+
A quick feature comparison as of June 2026:
|
|
77
|
+
|
|
78
|
+
| Capability | ToolFunnel | FastMCP | mcpproxy-go | MetaMCP | mcp-anything |
|
|
79
|
+
|---|:--:|:--:|:--:|:--:|:--:|
|
|
80
|
+
| Config-declared polyglot tools (no SDK) | ✓ | ✗ (decorators) | partial (proxy only) | ✗ | ✓ |
|
|
81
|
+
| Lean server-side exposure (briefs + schema on demand) | ✓ | partial | ✓ | ✗ | ✗ |
|
|
82
|
+
| Server-side fail-closed policy gate (Pre/Post) | ✓ | ✓ | ✓ | ✓ | ✗ |
|
|
83
|
+
| Live attach / hot-reload, no restart | ✓ | ✗ | partial | partial | ✓ |
|
|
84
|
+
| Non-Docker self-healing reconnect | ✓ | ✗ | partial (Docker) | partial (circuit-breaker) | ✗ |
|
|
85
|
+
| Visibility **matrix** - lean default + per-tool promote-to-every-turn + hide | ✓ | ✗ | filter only | cherry-pick only | ✗ |
|
|
86
|
+
| Zero runtime dependencies\* | ✓ (Node, no SDK) | ✗ (framework) | ✓ (Go) | ✗ (Docker) | ✓ (Go) |
|
|
87
|
+
| Runtime dependencies (installed / bundled) | **0\*** | many | 40+ (bundled) | many | bundled (Go) |
|
|
88
|
+
| Config web UI | ✓ | ✗ | ✓ | ✓ | ✗ |
|
|
89
|
+
|
|
90
|
+
Beyond the dependency count, four capability choices set ToolFunnel apart from a pure proxy like mcpproxy-go:
|
|
91
|
+
|
|
92
|
+
1. **It HOSTS, not just proxies.** A proxy forwards tools from existing MCP servers. ToolFunnel also turns an arbitrary local command or script, any language, into a first-class gated MCP tool from one JSON entry, with no SDK and no pre-existing server.
|
|
93
|
+
2. **Reference mode.** A tool ToolFunnel only *describes*: `toolfunnel_run_tool` hands back the instructions and the connected AI performs the action in its own environment - and the handoff itself is gated (a PreToolUse deny withholds the instructions). Nothing executes server-side.
|
|
94
|
+
3. **A portable policy gate, not a content scanner.** The gate speaks the Claude-Code hook protocol (PreToolUse/PostToolUse, exit-2-to-block) inside the gateway, so an existing hook/policy travels to any client unchanged - an easily adaptable (to CODEX or whatever you are using) programmable policy that ports, rather than a built-in scanner.
|
|
95
|
+
4. **The visibility matrix.** Not on/off filtering: a lean default you can selectively promote to the every-turn surface, hide from the manager view, or collapse entirely (turn the meta-tools off and present your tools as a plain top-level MCP).
|
|
96
|
+
|
|
97
|
+
Underneath it all: a lean surface (briefs + instructions-on-demand) that keeps token cost flat as you add tools, a fail-closed gate on every server-side path, and zero runtime dependencies\* in a component that sits in a privileged position. The wedge is the whole package in one small, auditable server, not any single trick.
|
|
98
|
+
|
|
99
|
+
### State of development and honest limitations
|
|
100
|
+
|
|
101
|
+
ToolFunnel is a focused solo build. I've tested it thoroughly for my own use case, but it may still contain bugs outside of that - and there are features I've leaned on less than others. The OAuth 2.1 and Streamable-HTTP implementations in particular are recent, added in line with the latest MCP SDK capabilities, and haven't been tested as extensively, so your mileage may vary.
|
|
102
|
+
|
|
103
|
+
There are no Prometheus/OpenTelemetry metrics either - instead there's a toggleable JSONL audit log plus in-memory call counters on `/health`. I haven't needed full metrics for my own use; if it's something you'd want, let me know and I'll look at adding it.
|
|
104
|
+
|
|
105
|
+
If you find a bug - better still, a bug and a fix - or have an improvement, I'd genuinely like to hear from you. I'm keen to work with others on ToolFunnel :)
|
|
106
|
+
|
|
107
|
+
And because it's hand-rolled rather than built on the MCP SDK, I'll need to keep it in line with MCP itself as the spec evolves - the revisions land roughly quarterly, and I aim to track new capabilities as they're announced. Again: if you'd like to get involved, I'd love to hear from you :)
|
|
108
|
+
|
|
109
|
+
## Use cases
|
|
110
|
+
|
|
111
|
+
- **Tame context bloat.** Several rich MCP servers connected at once drown the agent's context in tool schemas it won't use this turn. Behind ToolFunnel the agent sees short briefs and loads a tool's full schema only when it actually reaches for it.
|
|
112
|
+
- **One toolbox, many workflows.** Keep all your multi-use tools and common MCP servers in one place and pick which to surface per workflow, instead of re-wiring a different setup for every client.
|
|
113
|
+
- **Govern what an agent may run, on any client.** Put a fail-closed PreToolUse policy in the gateway so it travels with your tools, even to clients that have no hook system of their own.
|
|
114
|
+
- **Turn a script into an MCP tool without building an MCP.** Have a useful CLI, a bash one-liner, or a script in any language? Declare it in a JSON entry and it's a gated MCP tool - no protocol code, no SDK.
|
|
115
|
+
- **Keep a legacy internal MCP working.** Front an ageing in-house MCP server with ToolFunnel and keep the workflows that depend on it running as the ecosystem moves on.
|
|
116
|
+
- **Add a tool mid-session.** Need a tool while you're working? Drop it in and it's live on the next turn, no restart.
|
|
117
|
+
- **Survive an upstream crash.** If an attached MCP server dies mid-session, ToolFunnel notices, reconnects it in the background with backoff, and re-advertises its tools. Most reconnect logic elsewhere is Docker-scoped; this isn't.
|
|
118
|
+
|
|
119
|
+
## Zero dependencies, on purpose
|
|
120
|
+
|
|
121
|
+
ToolFunnel has no runtime npm dependencies (`"dependencies": {}`) and does not use an MCP SDK - the JSON-RPC 2.0 wire protocol is hand-rolled on top of Node built-ins (`node:http`, `node:child_process`, `node:fs`, …). Requires Node >= 18. `npm install toolfunnel` pulls zero packages.
|
|
122
|
+
|
|
123
|
+
> **\* The one asterisk: OAuth is opt-in.** Enabling OAuth 2.1 (off by default) adds exactly one dependency - [`jose`](https://www.npmjs.com/package/jose), which is itself zero-dependency, audited, and the same library the official MCP SDK uses. It is not a runtime dependency of the core: it appears only as a `devDependency` (for the OAuth test suite) and is installed on demand for users who turn auth on. So the default footprint is genuinely zero, and the most security-sensitive code in the project - token validation - is delegated to an audited library rather than hand-rolled.
|
|
124
|
+
|
|
125
|
+
This is a deliberate security decision, not minimalism for its own sake. ToolFunnel sits in a privileged position - it gates tool execution - which is exactly where you don't want an unaudited dependency tree. And `dependencies: {}` in npm is a stronger claim than a "single binary" elsewhere: a Go gateway that ships as one binary still statically links its dependency tree (mcpproxy-go's, for example, runs to 40+ direct deps - goja, esbuild, gRPC, an observability stack). That's dependency-bundled, not dependency-free. In a year of high-profile npm supply-chain compromises, a tool-execution gate with nothing transitive to audit is a defensible engineering stance - you audit your code, not forty supply chains. (It is not a silver bullet: a zero-dep posture shrinks the audit surface, it doesn't remove the burden of getting the hand-rolled wire + gate correct - which I have tried to do.)
|
|
126
|
+
|
|
127
|
+
## Authentication - optional OAuth 2.1
|
|
128
|
+
|
|
129
|
+
By default the gateway is loopback-only and unauthenticated - the right posture for a single operator on `localhost`. For a networked or multi-user deployment, ToolFunnel can act as an OAuth 2.1 resource server: it validates the bearer token on every request before any tool runs, through `jose` with a pinned algorithm allowlist, enforced issuer, and an enforced audience bound to this gateway's resource URI (the RFC 8707 confused-deputy defence). Auth is what unlocks a non-loopback bind: the HTTP host refuses to bind off-localhost unless OAuth is enabled. Setup, config fields, and discovery details: [manual](docs/MANUAL.pdf).
|
|
130
|
+
|
|
131
|
+
The OAuth client leg (Dynamic Client Registration, authorization-server-metadata discovery, step-up auth) and the 2026 spec hardening are planned, not yet shipped - the resource-server slice is the shipped, tested MVP.
|
|
132
|
+
|
|
133
|
+
## Quickstart
|
|
134
|
+
|
|
135
|
+
[](https://glama.ai/mcp/servers/Rendeverance/toolfunnel)
|
|
136
|
+
|
|
137
|
+
```bash
|
|
138
|
+
npm install toolfunnel # runtime dependencies: none
|
|
139
|
+
# or from source:
|
|
140
|
+
git clone https://github.com/Rendeverance/toolfunnel.git && cd toolfunnel && npm install
|
|
141
|
+
|
|
142
|
+
node bin/toolfunnel.js # stdio MCP server (what most clients spawn)
|
|
143
|
+
node bin/toolfunnel.js --ui # optional config web UI on 127.0.0.1:9777
|
|
144
|
+
```
|
|
145
|
+
|
|
146
|
+
Point your client at it via `.mcp.json`:
|
|
147
|
+
|
|
148
|
+
```json
|
|
149
|
+
{ "mcpServers": { "toolfunnel": { "command": "node", "args": ["/path/to/toolfunnel/bin/toolfunnel.js"] } } }
|
|
150
|
+
```
|
|
151
|
+
|
|
152
|
+
**Or skip the config entirely - just ask your AI.** Point any MCP-aware client at ToolFunnel and say *"attach this tool server, expose these tools, and block anything destructive."* The AI reads ToolFunnel's own built-in instructions and wires it all up for you - upstreams, exposed tools, and a PreToolUse safety gate - in plain language, no JSON editing. Prefer to click? The web UI does the same. No coding experience required :)
|
|
153
|
+
|
|
154
|
+
The HTTP host (`--http`), registering over HTTP, the web UI tabs, the audit log, configuration files, project layout, and the smoke test are all walked through step-by-step, with screenshots, in the [manual](docs/MANUAL.pdf).
|
|
155
|
+
|
|
156
|
+
## Packaging: ship your own MCP
|
|
157
|
+
|
|
158
|
+
Everything you build here - tools, curation, upstream selections, policy hooks, identity - lives in one folder (the config home), and one management call packages it:
|
|
159
|
+
|
|
160
|
+
- `tf_pack { format: "home" }` produces a portable setup: zip it, git it, or run it with `--config-dir`.
|
|
161
|
+
- `tf_pack { format: "npm", name: "my-mcp" }` produces a publishable npm package that depends on toolfunnel (never forks it), bundles your setup, and gives your users `npx my-mcp` - your server, your name in the handshake, your tools and schemas, your gate enforced on their machine regardless of client.
|
|
162
|
+
|
|
163
|
+
Full story: [docs/packaging.md](docs/packaging.md).
|
|
164
|
+
|
|
165
|
+
## License
|
|
166
|
+
|
|
167
|
+
[MIT](LICENSE)
|