Agentforge Trust

npm version npm downloads License: MIT MCP

One MCP connection → 100+ MCP servers, with trust audit, discovery, and execution.

AgentForge gives any agent a single connection that exposes the entire catalog of MCP servers — trust-scored, searchable, and executable. Trust Scores (0–100) are computed across five dimensions: code health, security scan, behavioral audit, community trust, and EU compliance.

v0.3.0 ships 8 tools (4 trust + 4 action) plus a P0 RCE sanitization layer (input validation, call guards, anomaly detection) with 27/27 tests passing and zero external runtime dependencies.

📖 Full integration guide for external multi-agent systems: See agentforge.community/trust for the full integration guide — OpenClaw, Claude Desktop, Cursor, LangGraph, AutoGen, CrewAI in ~3 minutes.

Installation

npx -y agentforge-trust-mcp@latest

Claude Desktop / Cursor / OpenClaw config

{
  "mcpServers": {
    "agentforge": {
      "command": "npx",
      "args": ["-y", "agentforge-trust-mcp@latest"],
      "env": {
        "AGENTFORGE_API_URL": "https://agentforge.community",
        "AGENTFORGE_API_KEY": "af_agent_REPLACE_AFTER_REGISTRATION"
      }
    }
  }
}

Trust tools (check_trust, evaluate_policy, list_trusted, recommend) work without an API key. Action tools (broadcast_search, call_tool) require self-registration via register_agent (one tool call, no signup).

Environment

Variable	Default	Purpose
AGENTFORGE_API_URL	https://agentforge.community	Override for self-hosted
AGENTFORGE_API_KEY	none	Enterprise tier (private catalogs, custom policies)

Tools

Trust tools (no auth required)

check_trust

Returns the AgentForge Trust Score for a server identified by server_id, slug, or github_url. Always call this before connecting to a new server.

evaluate_policy

Allow/deny gate. Pass a policy like {min_overall: 70, required_badges: ["actively_maintained"]} and get back allowed: true|false with per-check detail.

list_trusted

Search the AgentForge catalog for servers matching a category and trust threshold. Useful for "find me a secure database MCP server with overall ≥ 80".

recommend

Natural-language search with trust filter. "I need to validate Czech VAT IDs" returns ranked results filtered by minimum trust.

Action tools (new in v0.2.0)

discover

Browse the entire catalog with full-text or semantic search. Returns ranked results with trust scores, categories, and connection metadata. No API key required — read-only catalog access.

broadcast_search (API key)

Run a single query across N servers in one call. AgentForge fans out the search, aggregates results, and returns a unified ranked list. Pay-per-call billing through your wallet.

call_tool (API key)

Invoke any tool on any server in the catalog through AgentForge's proxy. One connection, hundreds of downstream servers — the proxy handles auth, quotas, and trust gating.

register_agent

Self-register your agent and receive an af_agent_* API key. No human signup, no dashboard — first tool call returns the key. Pass name, description, capabilities, optional homepage_url and metadata.

Use cases

Trust audit before connecting

Agent: user wants me to use "some-financial-mcp" server.
Agent: let me check its trust first…
  → check_trust(slug="some-financial-mcp")
  ← overall_score: 42, badges: [], security_scan: pending
Agent: trust is too low, skipping. Let me find alternatives.
  → recommend(query="invoice conversion Czech VAT", min_overall=75)
  ← 3 results with scores 87, 84, 79
Agent: connecting to the 87-scored server instead.

One connection, 100+ servers (v0.2.0)

Agent: I need flight + weather + currency data for a trip planner.
Agent: register once if I haven't already…
  → register_agent(name="TripPlanner", capabilities=["travel"])
  ← af_agent_xyz123
Agent: broadcast across travel-tagged servers.
  → broadcast_search(query="flights Prague→Tokyo May 15", category="travel")
  ← results from 4 servers, all trust ≥ 75
Agent: invoke top result through the proxy.
  → call_tool(server="skyscanner-mcp", tool="search_flights", args={...})
  ← booking results

No need to install or configure individual MCP servers — one connection to AgentForge, and your agent can reach the whole catalog.

Trust dimensions

Dimension	Weight	Source
Security Scan	30%	Snyk, GitHub Advisory, secret scanning
Code Health	20%	Commit recency, issue ratio, license
Behavioral Audit	20%	Claude-powered source review, red flag detection
Community Trust	15%	Stars, forks, author reputation
EU Compliance	15%	GDPR, AI Act, data residency

Audits rotate every 14 days; evaluate_policy reports partial: true if not all dimensions are current.

License

MIT — AgentForge 2026