Moltline Agent Governance

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Audit MCP configs and skill files for over-broad scope and injection risk. 6 of 8 free.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "Moltline Agent Governance": {
      "url": "https://mcp.moltlinestudio.com/govern",
      "transport": "streamable-http"
    }
  }
}

Measured tools (8)

• audit_mcp_config read-only

  Audit an MCP server config for risk-ranked posture findings. FREE. Flags exposed machine credentials in the config, required inputs that aren't gated/optional, unpinned versions, over-broad env access, and dangerous auto-run flags. It never echoes any matched secret value back. Typical input {"config": "<mcpize.yaml, mcp.json, or a Claude/Cursor servers block>"} returns {"posture_score": 0-100, "verdict": "...", "findings": [{"line": N, "severity": 1-5, "issue": "...", "fix": "..."}], "note": "..."}. Use on a server configuration document. Not for a skill or instruction file (audit_skill_file) and not for untrusted content an agent is about to read (injection_scan). Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • config (required) — string — The MCP config to audit, pasted as text or JSON (mcpize.yaml, mcp.json, or a Claude/Cursor servers block).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• scope_check read-only

  Score the blast radius of every tool in a permission manifest. FREE. Ranks each tool by capability risk (command exec > money/delete > file-write/messaging > read > network) and flags the over-privileged ones that need approval gates. Typical input {"tools": "[\"run_shell\", \"read_docs\"]"} returns {"tools_scored": N, "high_risk_tools": N, "ranking": [{"tool": ..., "blast_radius": 0-5, "capabilities": [...]}], "recommendation": ["..."], "note": "..."}. Use on a permission manifest to rank tools by blast radius. Not for the configuration that mounts them (audit_mcp_config). Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • tools (required) — string — The manifest as a string (a JSON array of tool names or {name, description} objects, a JSON object of name->description, or plain newline-separated names).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• audit_skill_file read-only

  Audit an agent skill or instruction file before you trust it. FREE. Checks for governance smells: prompt-injection and guardrail-bypass phrasing, concealment instructions ('don't tell the user'), exfiltration language, and exposed credential material. Typical input {"content": "<SKILL.md, system prompt, or tool description text>"} returns {"verdict": "reject: do not install" | "no governance red flags on a pattern pass", "findings": [{"severity": 1-5, "issue": "..."}], "note": "..."}. Use before trusting a skill or instruction file that came from outside your own repository. Not for arbitrary untrusted input at run time (injection_scan). Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • content (required) — string — Full text of the skill file, system prompt, or tool description to audit.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• injection_scan read-only

  Scan untrusted text for prompt-injection patterns before ingestion. FREE. Use on any web page, email, or document an agent is about to ingest to catch prompt-injection and data-exfiltration patterns before they reach the agent's context. Typical input {"text": "<untrusted content>"} returns {"injection_suspected": bool, "count": N, "hits": [{"line": N, "pattern": "...", "text": "<flagged line>"}], "note": "..."}. Not for reviewing a skill file you control (audit_skill_file), and a clean result is not a guarantee of safety - it reports pattern matches only. Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • text (required) — string — The untrusted content to scan, pasted as a single string.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• inventory_report read-only

  Build a governance inventory with risk tiers from a raw agent list. FREE. Turns a list of agents / MCP servers / skills into an audit-ready summary with critical/elevated/standard tiers and unowned-agent flags. Typical input {"items": "[{\"name\": \"deploy-bot\", \"owner\": \"ana\"}]"} returns {"total": N, "tiers": {"critical": N, ...}, "unowned_agents": [...], "inventory": [{"name": ..., "owner": ..., "tier": ..., "orphaned": bool}], "reading": "...", "note": "..."}. Use to turn a raw agent list into risk tiers. Not for auditing any single agent in depth (audit_mcp_config, scope_check). Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • items (required) — string — The fleet as a string (a JSON array of {name, owner?, capabilities?, last_seen?} objects, or plain newline-separated agent names).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• governance_policy read-only

  Generate an audit-ready agent-governance policy for a fleet. PREMIUM (license). Covers inventory cadence, ownership rules, least-privilege approval gates, injection defense, logging/retention, and decommissioning triggers. Typical input {"fleet_context": "20 agents, 3 with shell access, one finance bot"} returns {"policy": ..., "sections": {...}, "context_note": ..., "audit_checklist": ["...", ...]}. Use when a fleet needs a written policy document. Not for assessing what the fleet currently does (inventory_report, audit_mcp_config). Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • fleet_context — string — Optional plain-language description of the fleet (size, capabilities, sensitive systems) used to tailor the policy; empty returns the generic baseline.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_auditor_persona read-only

  Load the Governance Auditor persona for consistent fleet audits. PREMIUM (license). The persona is methodical, evidence-driven, and allergic to 'it's probably fine'. Takes no arguments. Returns {"persona": ..., "identity": ..., "rules": ["...", ...], "opening_move": "..."} ready to adopt as a system prompt. Use to keep repeated audits consistent in voice and rigor. Not for running an audit - the audit tools do that. Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"}. Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• agent_readiness_scan read-only

  Score a public domain against 21 agent-readiness checks. FREE. Use when you need to know whether an autonomous agent can discover, read, use or pay a website - your own, or a vendor you are evaluating before recommending it. Typical input {"domain": "example.com"} returns {"score": 8, "total": 21, "grade": "F", "passed": [...], "failed": [{"title": "...", "detail": "...", "fix": "..."}], "report_url": "..."} where report_url is a permanent shareable page for the same result. Not for auditing an MCP client configuration (audit_mcp_config) and not for scanning text for injection (injection_scan) - this one reaches out over the network and fetches public URLs on a live domain. Errors: on invalid, missing, or malformed input this tool never raises a protocol error; it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "The readiness scanner is not reachable right now."}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

  Parameters:

  • domain (required) — string — A public hostname such as example.com. A full URL is accepted and reduced to its host. Hostnames that resolve to private or internal addresses are refused.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

Publisher website Repository

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.