Jmunch Mcp

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.

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:

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 to reproduce.

Install

pip install jmunch-mcp

From source:

git clone https://github.com/jgravelle/jmunch-mcpcd jmunch-mcppip install -e .

Quickstart

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

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.

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

Point your app:

# 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.

jmunch-mcp dashboard              # http://127.0.0.1:7878jmunch-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.

License

jmunch-mcp is released under the MIT 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.