munio 0.2.0

[m] munio
Security scanner and runtime guard for AI agent tool calls

AI agents call external tools — MCP servers, OpenClaw skills, API endpoints. A malicious or poorly-written tool can exfiltrate your data, execute arbitrary commands, or chain actions into multi-step attacks. munio catches these issues before they reach your agent.

pipx install munio
munio config-scan

For deep tool schema analysis:

munio scan --server "npx @modelcontextprotocol/server-filesystem /tmp"

No MCP servers? Try the bundled example:

munio scan --file examples/vulnerable-server.json --details

Why scan MCP servers?

The same vulnerability classes that led to 512 findings in Copilot extensions and 820+ malicious OpenClaw skills -- path traversal, command injection, SSRF, prompt injection -- exist in every AI tool-calling ecosystem. MCP servers, OpenClaw skills, and framework-integrated tools share the same attack surface.

munio was built by scanning 700+ public MCP servers and responsibly disclosing the vulnerabilities found. It works with MCP, OpenClaw, LangChain, CrewAI, and OpenAI Agents SDK.

What it catches

Installation

pipx install munio          # CLI (recommended)
pip install munio            # library
pip install "munio[z3]"      # with formal verification
pip install "munio[all]"     # everything

Scan MCP servers

munio scan --server "npx @foo/mcp-server"       # scan a live server
munio scan --file tools.json                     # scan exported schemas
munio scan                                       # auto-discover from IDE configs
munio scan --details                             # show affected tools and fixes
munio scan --format sarif --output report.sarif  # SARIF 2.1.0 for CI

8 analysis layers: L1 Schema, L2 Heuristic, L2.5 ML Classifier, L2.6 Multilingual ML, L3 Static, L4 Z3 Formal, L5 Compositional, L7 Source.

Scan config files

Finds hardcoded credentials, unpinned dependencies, and supply chain risks in Claude Desktop, Cursor, VS Code, Windsurf, Cline, and Junie configs. No server connections needed.

munio config-scan                    # auto-discover all IDE configs
munio config-scan --config file.json # scan a specific config

Detect cross-server attack chains

munio compose --schemas-dir ./schemas  # analyze pre-fetched schemas
munio compose --format markdown        # generate CVE filing drafts

Protect at runtime

Intercept every tool call before execution. No code changes.

munio init      # wrap all MCP servers in IDE configs
munio status    # check protection status
munio restore   # remove wrapper

After munio init, every tools/call is verified against YAML constraints. Dangerous calls are blocked before reaching the server.

Constraint example

name: block-dangerous-urls
action: http_request
check:
  type: denylist
  field: url
  values: ["evil.com", "169.254.169.254"]
  match: contains
on_violation: block
severity: critical

8 check types: denylist, allowlist, threshold, regex_deny, regex_allow, composite, rate_limit, sequence_deny.

Python API

from munio import Guard

guard = Guard(constraints="generic")
result = guard.check({"tool": "http_request", "args": {"url": "https://evil.com"}})
# result.allowed = False

Adapters for LangChain, CrewAI, OpenAI Agents SDK, and MCP. See docs.

How it works

The runtime gate is 100% in-process and deterministic — no SMT solver, no subprocess, and the [z3] extra is not required to run it:

For a constraint like cost * quantity <= budget, munio decides whether any in-bounds value of a missing variable could violate it by evaluating the polynomial at the bounds — an exact answer, no solver. Constraints outside that grammar are rejected at load time with an actionable message, never mis-decided.

Z3 is used only offline, where satisfiability is the real problem and there's no latency pressure: deploy-time policy verification (munio policy) and static scanning (munio scan layer L4). Both are in the optional [z3] extra.

All commands

Development

git clone https://github.com/munio-dev/munio.git && cd munio
make install    # uv sync + pre-commit hooks
make test       # 3900+ tests
make ci         # lint + typecheck + tests + coverage

License

Apache 2.0