Skip to main content
Glama
README.md
# jmunch-mcp

<!-- mcp-name: io.github.jgravelle/jmunch-mcp -->

Transparent MCP proxy that reduces the token cost of large upstream tool responses for nearly every other MCP server imaginable. Wraps a single upstream MCP, forwards every call, and handle-ifies fat payloads into content-aware backends the agent can query with a small set of universal verbs (`peek`, `slice`, `search`, `aggregate`, `describe`, `list_handles`).

## Benchmarks

Measured end-to-end against two popular real-world MCP servers. Each run fires a fixed script of tool calls twice — once direct, once through jmunch-mcp — with three follow-up `jmunch.*` verb calls on the proxied side to model an agent drilling into a large result rather than slurping it whole.

| suite | upstream | direct tokens | via jmunch-mcp | saved |
|---|---|---:|---:|---:|
| GitHub (`facebook/react` issues/PRs/commits) | `@modelcontextprotocol/server-github` | 379,878 | 44,328 | **335,550 (88.3%)** |
| Firecrawl (Wikipedia scrapes + site map + search) | `firecrawl-mcp` | 259,574 | 2,928 | **256,646 (98.9%)** |

Wall-clock time was also faster with the proxy on both suites, despite the extra verb calls — the agent never has to page through the fat payload:

| suite | direct | via jmunch-mcp | delta |
|---|---:|---:|---:|
| GitHub    |  8.4s |  6.8s | **−1.6s (−19.0%)** |
| Firecrawl | 16.4s |  9.2s | **−7.2s (−43.9%)** |

Tabular content (GitHub) routes to the SQLite backend and answers `peek`/`slice`/`aggregate`; JSON content (Firecrawl scrape/map) routes to the JSON-tree backend and answers `peek`/`slice` (JSONPath)/`search`. See [bench/README.md](bench/README.md) to reproduce.

## Install

```bash
pip install jmunch-mcp
```

From source:

```bash
git clone https://github.com/jgravelle/jmunch-mcp
cd jmunch-mcp
pip install -e .
```

## Quickstart

```bash
jmunch-mcp init
```

`init` scans three sources — your MCP client configs (Claude Desktop, Claude Code, Cursor, Windsurf, Continue), running processes, and a small catalog of popular upstreams (GitHub, Firecrawl, filesystem, fetch, Brave Search, Slack) — and renders a checklist. Tick the upstreams you want wrapped, and it writes one `<name>.toml` per selection into `./configs/`. Non-interactive flags: `--yes` (pick everything already registered in a client), `--dry-run`, `--overwrite`, `--out <dir>`, `--no-running`, `--no-catalog`.

### Manual

```bash
jmunch-mcp --config examples/config.toml
```

Configure your MCP client to launch `jmunch-mcp --config <path>` instead of the upstream server directly. Add `--report` to print a session summary on shutdown.

## Gateway mode (v2 — universal proxy)

The MCP proxy above saves tokens for MCP clients. The **gateway** saves tokens for *any* AI application that speaks the OpenAI or Anthropic HTTP API — LangChain, LlamaIndex, CrewAI, AutoGen, Continue, Cline, Aider, or a raw SDK. No code changes in the app; just point `base_url` at jmunch.

```bash
pip install 'jmunch-mcp[gateway]'
jmunch-mcp gateway --config configs/gateway.example.toml
# listening on http://127.0.0.1:7879
```

Point your app:

```bash
# OpenAI SDK, LangChain, Aider, Continue, Cline, Ollama-compat apps:
export OPENAI_API_BASE=http://127.0.0.1:7879/v1

# Native Anthropic SDK / Claude Code:
export ANTHROPIC_BASE_URL=http://127.0.0.1:7879
```

What it does, transparently:

- **Handle-ifies fat tool_results** in outgoing requests — your app's tool returns 100KB of JSON, the model sees a 1KB summary + opaque handle.
- **Injects jmunch verbs** (`peek`, `slice`, `search`, `aggregate`, `describe`, `summarize`, `list_handles`) into the request's `tools` array so the model can drill in.
- **Short-circuits verb calls** — when the model calls `jmunch_peek`, the gateway resolves it locally against the handle registry and synthesizes the follow-up turn. The app never sees jmunch tool_calls; those completions cost zero upstream tokens.
- **Persists handles** to `~/.jmunch/handles.db` with a configurable TTL so they survive restarts and cross-session reads.
- **Streams both ways** — OpenAI SSE and Anthropic event streams are buffer-then-replayed with correct verb resolution.

Per-request controls via headers:

- `X-Jmunch-Upstream: <name>` — override the configured upstream.
- `X-Jmunch-Inject: false` — disable tool injection for this call (pure pass-through + request-side handle-ify only).

Metrics flow into the same dashboard as the MCP proxy. Filter with `?surface=gateway` or `?surface=mcp` on `/api/stats` and `/api/calls`.

## Dashboard

A read-only local web UI over the metrics DB each proxy writes to. Shows cumulative totals, per-upstream breakdowns, and a time series of forwarded calls.

```bash
jmunch-mcp dashboard              # http://127.0.0.1:7878
jmunch-mcp dashboard --open       # also open in your default browser
```

Flags: `--port` (default `7878`), `--host` (default `127.0.0.1`), `--db <path>` to point at a non-default metrics DB, `--open` to launch the browser. Metrics only populate once proxies have recorded calls, so run your client against a wrapped upstream first.

## When the proxy exits

The proxy ends when its client closes stdin, when its upstream exits, or when the process that started it is gone. It then closes the upstream's stdin and, if the upstream has not exited 3 seconds later, terminates it, then kills it 3 seconds after that. This reaches the process the proxy started; an upstream that is itself a wrapper (`npx.cmd` runs under `cmd.exe` on Windows) can leave its own child running.

For the last case a background thread checks the parent process every 5 seconds. It reads only the process table: nothing is written and nothing goes over the network. A parent it cannot check is never treated as gone. Set `JMUNCH_PARENT_WATCH_SECONDS` to change the interval, or to `0` to turn the check off.

## License

jmunch-mcp is released under the [MIT License](LICENSE) — free to use, modify, distribute, and embed in commercial products.

Note that licensing of **upstream MCP services** you proxy through jmunch-mcp is governed by those services' own terms. This applies to third-party MCP servers (GitHub, filesystem providers, vendor APIs) and to any sibling tools in the broader retrieval ecosystem you may compose with — check each upstream's license before redistribution.