bastionsupply

MCP supply-chain security scanner. Point it at an MCP server's tool definitions and it flags the supply-chain attacks before you install: tool-poisoning, tool-shadowing, hidden unicode, secret solicitation, dangerous capabilities, and rug-pull drift.

The pre-flight leg of the bastion family:

Offline scanning is pure static analysis of what a server claims about itself — no LLM, no network. Its one dependency is bastioncorpus, the shared prompt-injection dataset the whole bastion family uses for attack signatures.

Install

pip install bastionsupply

Use

# offline: scan a tools/list JSON dump
bastionsupply scan tools.json

# live: spawn a stdio MCP server and scan the tools it advertises
bastionsupply scan --stdio "npx -y @some/mcp-server" --live

# live: scan a remote MCP Streamable-HTTP server
bastionsupply scan --http https://some.host/mcp --live

# live: scan every server in an MCP client config
bastionsupply scan --config ~/.config/mcp.json --live

# rug-pull: pin tool hashes now, detect silent changes later
bastionsupply lock tools.json -o supply.lock
bastionsupply verify tools.json --lock supply.lock

# bridge: emit an agentbastion tool policy (default-deny, risky tools blocked)
bastionsupply harden tools.json -o policy.yaml

scan exits non-zero when anything critical or high is found — drop it in CI to fail a build that pulls in a poisoned server.

What it catches

Library

from bastionsupply import load_json_file, scan, to_text, to_policy_yaml

report = scan(load_json_file("tools.json"))
print(to_text(report))
print("safe" if report.ok else "risky", report.risk)

Live fetch note

--stdio / --config --live spawn the server process to call tools/list. Only run them on servers you intend to execute. Offline scan tools.json never runs anything. Live fetch is bounded — a hostile server that hangs or streams a giant reply is cut off by the timeout and a per-message size cap, not left to hang or exhaust memory. --http speaks http/https only and never follows redirects (no SSRF to file:// or internal hosts), and reads multi-event SSE / batch replies.

Optional: semantic poisoning tier

Catch paraphrased injections the literal checks miss by embedding tool text and comparing it to bastioncorpus's malicious intents. Off unless an embedder is set:

pip install "bastionsupply[semantic]"
export BASTIONSUPPLY_EMBED_MODEL=all-MiniLM-L6-v2   # local, no egress
bastionsupply scan tools.json                        # now also runs semantic-poisoning

harden → graded policy

harden denies critical/high tools and, for allowed-but-sensitive tools (e.g. a network capability), emits graded caution the whole family understands: a rate_limits hint for agentbastion and a per-tool scrub_results override for bastiongate.

MIT.