mcpgawk

[mcpgawk by nativerse]

One gateway in the path. On your machine.

mcpgawk

[PyPI] [Python] [License] [CI] [Open VSX] [GitHub Marketplace] [No egress]

Your agents call Model Context Protocol servers that can change what their tools do after you approved them, and the agent will call the new one without noticing. mcpgawk reads every server your agents can reach, checks every call against a baseline you approved, and blocks the ones that changed. It runs on your machine and uploads nothing.

The same engine powers mcpgawk Platform: mcpgawk enforce puts one endpoint in front of the whole fleet with a key per caller, policy on every call and a hash-chained audit log, and mcpgawk monitor watches the servers you approved around the clock and tells you when one drifts. This free layer is the seeing and the blocking underneath it. mcpgawk Platform is one subscription per person for up to 3 machines: start a free 7-day trial at https://mcp.gawk.dev/trial.html (no card) or subscribe at https://mcp.gawk.dev/subscribe. Then mcpgawk login <key> on this same install fetches the paid engine and turns those on — one more command, nothing else to set up.

[mcpgawk in 64 seconds: a tool you approved changes in an update, and mcpgawk refuses the call until you decide. Watch the video.]

_{Watch: mcpgawk in 64 seconds. A tool you approved changes in an update; mcpgawk refuses the call until you decide.}

[mcpgawk demo: a server is approved, changes afterwards, and the guard blocks the tool that appeared]

_{mcpgawk demo — its own output, in a sandbox that touches nothing of yours. Run the same command and you get the same thing.}

Why

A server you approved can change what its tools do afterwards. Nothing in MCP tells your agent that happened — it just calls the new tool. That is the rug-pull, and it is the case mcpgawk is built for.

Two things follow from being able to see a server properly. You find out what each one can reach before you trust it, and you find out what it costs: every tool is loaded into your context on every request, used or not.

How it's different

• It blocks, it does not only report. A scanner tells you afterwards. mcpgawk guard installs one pre-execution hook and a tool that appeared after you approved the server does not run.
• It runs the server, not just reads it. mcpgawk verify drives tools in a sandbox and reports what they actually did — exfiltration, SSRF, poisoning — reproduced before it is reported.
• Nothing is uploaded. Cloud scanners send your inventory to a server and gate the verdict there. Every decision here is made on your machine, with no account and nothing to sign in to.
• It says what it did not check. Skipped tools are named as skipped, never counted as clean.

Features

• 🛑 Block a changed tool before it runs — mcpgawk guard install puts one pre-execution hook in your agent's loop. The decision is local, in about 10ms, with nothing to sign in to. Works on 6 of the 21 supported clients; the rest have no hook point and are named, not glossed over.
• 🧪 Run it, don't just read it — mcpgawk verify drives tools in a sandbox and reports what they did: exfiltration, SSRF, tool poisoning, secret leaks. The sandbox is a proxy by default, so it needs no Docker; Docker adds full container isolation when you have it. Safe mode drives only provably read-only tools, and every tool it skips is named as skipped.
• 🧑‍⚖️ Approval needs a person — mcpgawk decide opens a local screen for what changed. The buttons live on the tokened link printed in your terminal, so an agent that opened the page cannot approve its own way past a block.
• 🖥️ One local panel — mcpgawk panel: every server, every decision, every piece of evidence.
• 📜 The changelog no vendor publishes — mcpgawk changes shows every change to a server's tool surface between the snapshots you have recorded: tools added or removed, input schemas widened, descriptions and annotations rewritten. It reads your local history, so it works for a server you approved weeks ago — the one thing a fresh scan can never tell you.
• 🔌 Any transport — stdio, streamable-HTTP, SSE, and OAuth remotes (via the mcp-remote bridge).
• 💸 Token cost index — exactly what each tool adds to your context at connect, plus the 3 heaviest tools.
• 🧾 Capability facts — write / exfil-capable / declared annotations, straight from the schema, plus a trust-surface summary (% write, % exfil-capable, destructive-declared count) and an annotation-completeness score.
• 📌 Integrity pin + drift — catch a server that silently rewrites its tools (--track).
• 🚩 Bounded signals — injection-shaped descriptions, cross-server shadowing, under-declaring Server Cards — pointers for a human, never verdicts.
• 🔒 Zero egress, by construction — the measurement layers import no network library. Enforced by a test. Two checks are opt-in and make an explicit exception (see Guarantees): --supply-chain and --oauth-scopes.

Get it — three ways

CLI (any terminal):

uv tool install --force mcpgawk      # or: pipx install --force mcpgawkmcpgawk                              # finds every agent config on the machine itself

Run it as an MCP server: mcpgawk mcp (stdio), or uvx mcpgawk mcp.

Editor (VS Code / Cursor): install mcpgawk from the marketplace (Open VSX). It scans your workspace mcp.json and shows cost + capability flags inline. The extension drives this engine as a subprocess — it is built and released separately, so its source is not in this repository.

CI (GitHub Action): gate every PR on token budget / drift (Marketplace):

- uses: gawk-dev/mcpgawk@v1  with: { config: mcp.json, max-tokens: 8000, fail-on-flagged: true }

When to run it

• Once, then leave it on — mcpgawk guard install. After that a tool that appears on a server you already approved does not get called.
• Before you add a server — see what it costs and what it can do, before you trust it.
• When your agent feels slow or picks the wrong tool — it's often MCP bloat (too many / too-heavy tools).
• On every PR — the CI gate catches drift and creeping token cost.
• If you publish an MCP server — see what it costs your users and how it reads to a client, and fix it (usually one line per tool). Lean + well-annotated is a differentiator.

Use

mcpgawk                                  # first run: every agent config on this machinemcpgawk demo                             # the whole arc in a sandbox — approve, drift, blockmcpgawk guard install                    # put the baseline in your agent's loopmcpgawk guard status                     # is protection actually on?mcpgawk decide                           # what changed, and approve it as a humanmcpgawk panel                            # the local page: servers, decisions, evidencemcpgawk verify mcp.json                  # run the servers and watch what they domcpgawk changes                          # what changed on your servers since you approved themmcpgawk login <key>                      # licence? one command fetches the paid engine (enforce, monitor)

Scanning on its own, if that is all you want:

mcpgawk scan mcp.json                                              # a whole configmcpgawk scan --stdio "npx -y @modelcontextprotocol/[email protected] /tmp"mcpgawk scan --http https://host/mcp --header "Authorization: Bearer $TOKEN"mcpgawk scan --sse  https://host/ssemcpgawk scan mcp.json --track                                     # record + detect rug-pulls over timemcpgawk scan mcp.json --json                                      # machine-readable labelsmcpgawk scan mcp.json --verbose                                   # full per-tool table, not just flagged toolsmcpgawk scan mcp.json --supply-chain                              # opt-in: npm/PyPI deprecation check (network)mcpgawk scan mcp.json --oauth-scopes                              # opt-in: decode a supplied Bearer JWT's scope

What it reports

• Cost index — tokens each tool adds at connect (named tokenizer; a comparable index, not an absolute Claude count), plus the 3 heaviest tools.
• Trust surface — capability facts (write/mutating, exfil-capable, declared annotations) rolled up into % write, % exfil-capable, and a destructive-declared count.
• Annotation completeness — a transparent composite (annotated ÷ total tools declaring read/write intent), not a risk score.
• Coverage — tools, prompts, and resources counted (--verbose for the full per-tool table).
• Integrity pin — a hash that changes if the server silently rewrites its tools; --track turns it into rug-pull detection over time.
• Bounded signals — precise, low-false-positive pointers for a human to review, never verdicts: injection-shaped descriptions (tools and prompts), cross-server name shadowing, and public Server Cards that under-declare what the server actually exposes.
• Supply-chain (opt-in, --supply-chain) — checks the launched package against the public npm/PyPI registry for deprecation/yank status.
• OAuth scopes (opt-in, --oauth-scopes) — locally decodes a supplied Bearer JWT's scope claim.

Guarantees

• No inventory egress. The only network is the protocol client talking to the server you point it at. The measurement layers import no network library — they cannot egress by construction (enforced by a test). Public Server Card discovery is fetched with no auth and no redirect-following. Two flags are the explicit, opt-in exception: --supply-chain sends the launched package's name (and pinned version, if any) — never your tool inventory — to the public npm registry or PyPI JSON API. --oauth-scopes makes no network call at all; it locally decodes a Bearer JWT you already supplied. Neither runs unless you pass the flag.
• Facts ≠ heuristics. Exact capability facts and the token index never mix with the bounded heuristic signals — separate in code, separate in output.
• Reproducible. One command, identical numbers.
• Tracks the protocol. Built on the official mcp SDK, which negotiates the protocol version.

Develop

uv run --extra dev --with mcp --with tiktoken --with httpx python -m pytest -q

CI gate — GitHub Action

Scan your MCP servers on every pull request and fail the build if one gets too heavy or trips a signal. It runs entirely in your runner — nothing is uploaded — and posts a per-server cost/flag table to the job summary.

- uses: gawk-dev/mcpgawk@v1  with:    config: mcp.json        # or: stdio / http / sse — a single server    max-tokens: 8000        # fail if any server loads more than this at connect    fail-on-flagged: true   # fail if any bounded signal fires

Available on the GitHub Marketplace.

Contributing

Issues and PRs welcome. Please read CONTRIBUTING.md first, and see the design boundaries in THREAT-MODEL.md. Security reports go through SECURITY.md (privately, not a public issue).

License

Apache-2.0 — see LICENSE. Part of the nativerse · gawk.dev family. Site and docs: mcp.gawk.dev. The value is in the repo, not a cloud.

Use it from your agent (skill)

Let your coding agent run the checks itself — whenever it adds, upgrades or audits an MCP server:

# Claude Code (similar for other agents: copy the folder into their skills directory)mkdir -p ~/.claude/skills && cp -r skills/mcpgawk ~/.claude/skills/mcpgawk

The skill teaches the agent to measure a server BEFORE trusting it, audit an MCP-2 upgrade as a baseline diff instead of blind re-trust, and relay every consent prompt to you verbatim.