mcp-behave

Runtime behavioral auditor for MCP servers. Runs an MCP server under strace, exercises every tool / resource / prompt it advertises, and then compares what it declared against what it actually did — file reads, network egress, subprocess execution.

Static scanners read tool descriptions. mcp-behave watches behavior.

The idea in one contrast

targets/leaky_server.py and targets/honest_server.py advertise the identical tool:

format_note — "Formats a markdown note. Purely local text formatting."

A static scanner reads two harmless tools. mcp-behave sees the difference:

The honest server producing zero findings matters as much as the leaky one tripping two — false positives would kill credibility.

Install

pip install mcp-behave

Requires Linux + strace for the syscall-trace backend. On macOS/Windows, use the bundled Dockerfile (see below).

Use

# Audit any stdio MCP server -- pass its launch command:
mcp-behave python -m mcp_server_fetch
mcp-behave npx -y @modelcontextprotocol/server-filesystem /tmp
mcp-behave uvx mcp-server-git

# Machine-readable output for CI:
mcp-behave --json --fail-on high python -m mcp_server_fetch

# Connect to an already-running remote server (manifest analysis only,
# no syscall trace possible):
mcp-behave --transport sse http://localhost:8000/sse

Exit codes: 0 clean, 3 findings at or above --fail-on threshold, 2 strace/ptrace setup failed, 1 other error.

Useful flags

• --json — emit findings as JSON on stdout.
• --fail-on {never,info,high} — severity that triggers exit code 3 (default high).
• --no-dns — skip reverse-DNS lookups (faster, offline-safe).
• --timeout SECONDS — per-call timeout (default 15).
• --transport {stdio,sse,streamable-http} — remote transports skip syscall tracing.
• --out-dir PATH — where manifest.json and trace.log are written.

Push results to a dashboard

mcp-behave push uploads a JSON audit result to an mcp-behave dashboard.

# 1. Sign in at https://<your-dashboard>.vercel.app/settings and mint a token.
export MCP_BEHAVE_DASHBOARD_URL=https://<your-dashboard>.vercel.app
export MCP_BEHAVE_TOKEN=mcpb_...

# 2. Audit something with --json and pipe it to push.
mcp-behave --json python -m mcp_server_fetch \
  | mcp-behave push --server-name server-fetch --category network

# Or save to a file first, then push.
mcp-behave --json python -m mcp_server_fetch > audit.json
mcp-behave push audit.json --server-name server-fetch --category network

Flags worth knowing:

• --server-name NAME — canonical name on the dashboard. Defaults to the joined server command from the audit JSON; usually you want to set this.
• --description TEXT — used the first time this server is submitted.
• --github-url URL, --npm-package PKG, --author NAME — optional server metadata.
• --category {local,network,database,other} — defaults to other.
• --dry-run — print the JSON payload that would be POSTed and exit.

The endpoint is rate-limited per token (20 req/min); push exits 1 on HTTP error and prints the response body to stderr.

What it detects

• Undeclared network egress — IP:port destinations reached during tool calls (reverse-DNS resolved in findings when available).
• Sensitive-path reads — ~/.ssh/*, ~/.aws/*, ~/.env, ~/.netrc, etc.
• Rug-pull — the declared manifest hash changes between runs of the same server (e.g. a tool's description silently changes after the user trusts it). Hashes persist under $XDG_DATA_HOME/mcp-behave/manifests/.

Findings are framed as observations ("does X, undeclared"), never accusations.

Docker (works on macOS / Windows hosts)

docker build -t mcp-behave .
# default: runs the bundled leaky target so you can see findings immediately
docker run --rm --cap-add=SYS_PTRACE mcp-behave
# point at a real server
docker run --rm --cap-add=SYS_PTRACE mcp-behave python -m mcp_server_fetch

--cap-add=SYS_PTRACE is required for strace to attach inside the container.

The image bundles Node.js (for npx servers), uv (for uvx servers), and sandbox canary files (sandbox_home/) mounted as $HOME so credential reads are detectable.

Development

pip install -e ".[test]"
pytest -v

CI runs on Ubuntu (strace is Linux-only); see .github/workflows/ci.yml.

Known limits

• Linux-only ground truth via strace. eBPF/seccomp backend is on the roadmap — see ROADMAP.md.
• stdio is the only transport with syscall tracing. sse and streamable-http connect to a remote server we cannot trace; only manifest analysis and rug-pull detection apply there.
• Input synthesis is heuristic — JSON Schema format, key-name hints, and type defaults. Schema-coverage fuzzing (hypothesis-jsonschema) is roadmap.
• AF_UNIX egress isn't matched — a server exfiltrating over a unix domain socket would slip past the current network detection.
• Single call per capability — servers that misbehave only on specific inputs or after N calls may not be triggered.

License

MIT. See LICENSE.