mcp-gauntlet 0.3.0

An agentic evaluation harness for MCP servers — can an AI agent actually accomplish real tasks with this server's tools?

mcp-gauntlet

An agentic evaluation harness for MCP servers. Point it at any Model Context Protocol server and it answers the question the static analyzers don't: can an AI agent actually accomplish real tasks using this server's tools?

Why

The existing MCP quality tools (mcp-lighthouse, mcp-scorecard, mcp-checkup) are all static — they inspect schemas, count tokens, and lint descriptions. None of them run an LLM agent against the server to see whether it can actually complete tasks. That dynamic, agent-in-the-loop evaluation — with a real task-success rate, not just a conformance check — is what mcp-gauntlet does. The same live run also scans the tools' actual outputs for prompt-injection, catching tool-poisoning that a static description scan can't (a server that looks clean at list-time but poisons at call-time).

Google Lighthouse tells you your web page is well-formed. mcp-gauntlet tells you your MCP server is usable by an agent — with a task-success rate to prove it.

What it scores

Each run produces a graded report card (JSON + Markdown) across:

• Schema Health — valid JSON schemas, typed and described parameters.
• Description Quality — can an agent tell when and how to use each tool?
• Security Signals — a static scan for tool-poisoning / prompt-injection markers and hidden characters, covering the server's init instructions, its tool descriptions, and every string in each tool's input schema — titles, enums, defaults, examples, $defs entries and unknown extension keywords included. The whole schema is serialized into the model's prompt, so any string in it can carry a payload; scanning only description fields at the top level is trivially evaded by nesting one behind a $ref. A critical finding caps the overall grade.
• Agent Task Success — a live LLM agent attempts generated tasks using only the server's tools; LLM-judged and repeated for a success rate.
• Tool-Selection Accuracy — did the agent call the tools it was expected to?
• Tool Reliability — did the server's tools execute without error?
• Response Safety — a dynamic scan of the tools' live outputs for the same injection / poisoning markers, catching a server that looks clean at list-time but poisons at call-time. Reported (and it lowers the score) but doesn't cap on its own, since a fetch/filesystem server may faithfully pass through untrusted content.
• Robustness — does the server reject malformed input gracefully? A tool that publishes no argument schema at all scores zero here rather than being skipped: a server that declares no contract can't reject anything, and skipping it would make omitting schemas a way to score higher.

Leaderboard

A live leaderboard ranks popular public MCP servers by their gauntlet score: ghalebdweikat.github.io/mcp-gauntlet

Generate one yourself across any set of servers listed in a JSON file:

uv run mcp-gauntlet leaderboard --servers leaderboard.servers.json --out docs

Each server's raw result is saved to servers/<name>.json alongside its page, so the site can be rebuilt for free after a presentation change — no re-running (or re-paying for) the evaluation:

uv run mcp-gauntlet leaderboard --render-only --out docs

Only servers the agent actually scored share the ranked table. The overall is a weighted mean over the dimensions present, so a server the agent never ran against — no LLM configured, the backend rate-limited it, or every tool was excluded as possibly-mutating — skips Agent Task Success (the heaviest dimension) and would score systematically higher on a smaller denominator. Those are listed separately under Partially evaluated, each with the reason it wasn't ranked, rather than mixed in where an untested server could outrank a tested one.

Quickstart

uv sync --extra dev

# Static + robustness checks only — no API key required
uv run mcp-gauntlet run "python -m mcp_gauntlet.fixtures.good_server" --no-agentic

# Full gauntlet, including the live agent (Groq's free tier works)
echo "GROQ_API_KEY=gsk_..." > .env
uv run mcp-gauntlet run "npx -y @modelcontextprotocol/server-everything"

The LLM backend is provider-agnostic — any OpenAI-compatible endpoint (Groq by default; also OpenRouter, Together, or a local Ollama / vLLM). Runs are read-only by default: tools that look mutating (by name/description or a self-declared MCP destructiveHint) are excluded unless you pass --allow-writes. That exclusion is a best-effort heuristic, not a guarantee — pair it with read-only credentials or a throwaway environment for untrusted servers. Generated task sets are cached so scores are reproducible across runs.

Bundled good / bad fixture servers make it easy to see the difference:

uv run mcp-gauntlet run "python -m mcp_gauntlet.fixtures.bad_server"   # capped C — tool poisoning
uv run mcp-gauntlet run "python -m mcp_gauntlet.fixtures.good_server"  # A

With an LLM key, the bad fixture also trips Response Safety: its status_report tool has a clean description but poisons its output, so only the runtime scan catches it — the static description scan can't.

A server that hangs can't stall the run: every tool call is bounded by --tool-timeout (default 60s) and recorded as a failed call against the server's Tool Reliability, and --timeout (default 900s, 0 disables) caps the evaluation as a whole so a server that hangs during connect or tools/list still can't wedge the CLI. Raise --tool-timeout if a server is legitimately slow rather than stuck — the report says so when the limit is what stopped it.

Configuration

Configure via a .env file (copy .env.example and fill it in) or real environment variables:

Variable	Purpose
GROQ_API_KEY / GEMINI_API_KEY / OPENAI_API_KEY / OPENROUTER_API_KEY	API key for the provider the agent should use (only one needed). A free Groq key: console.groq.com/keys.
MCP_GAUNTLET_PROVIDER	Which provider: groq (default), gemini, openai, openrouter, or ollama (local).
MCP_GAUNTLET_MODEL	Model override for that provider (e.g. gemini-flash-latest). Defaults to a sensible per-provider model.

The --provider / --model CLI flags override these, and --base-url points at any OpenAI-compatible endpoint — a local Ollama / vLLM / LM Studio or a gateway:

uv run mcp-gauntlet run "npx -y @scope/pkg" --base-url http://localhost:11434/v1 --model llama3.1

No API key? Static mode

Everything except the live agent runs without an LLM. mcp-gauntlet run <server> with no key configured reports a static grade from the LLM-free checks — schema health, description quality, security signals, and robustness probes:

uv run mcp-gauntlet run "npx -y @modelcontextprotocol/server-everything" --no-agentic

Add --no-probe for a pure inspection that never executes any of the server's tools. The leaderboard behaves the same way — with no key it ranks servers on the static + robustness checks alone.

Use it in CI

Gate your MCP server's pull requests on its gauntlet score. Copy examples/gauntlet-ci.yml into your server's repo as .github/workflows/gauntlet.yml, point it at your server, and the build fails when the score drops below your threshold:

- name: Run the gauntlet
  run: uvx mcp-gauntlet@0.3.0 run "python -m your_server" --no-agentic --fail-under 60

Pin the version, as above: an unpinned uvx mcp-gauntlet would let a new release move your gate without a commit.

The static + robustness checks need no API key. To include the live agent evaluation, add an LLM key (e.g. GROQ_API_KEY) as a repository secret and drop --no-agentic. The report is uploaded as a build artifact.

License

MIT © Ghaleb Dweikat