agent-tools (x402 + MCP + A2A directory)

mcpserver — agent-tools.cloud directory + MCP discovery

The server that powers agent-tools.cloud — an open directory of x402 paid services with a free MCP discovery endpoint. Also listed on Smithery.

Note (2026-05-25): the previously hosted paid Qwen3.6 relay and the paid vertical endpoints (signal / onchain / defi / portfolio) were retired after 30 days of zero external conversions. This repository now serves the directory site and the free MCP discovery server only.

What this serves today

| Path | Auth | Purpose | |---|---|---| | GET / (host = agent-tools.cloud) | free | Directory site (search / browse) | | GET /api/v1/search | free | Faceted service search | | POST /api/v1/ask | free, rate-limited | LLM-ranked recommendations grounded in directory candidates | | GET /api/v1/services/{slug} | free | Service card with payment, call and quality metadata | | GET /api/v1/categories, /api/v1/stats | free | Directory facets and aggregate stats | | POST /api/v1/submit | free, rate-limited | Pending service submission with dedupe | | POST /mcp-discovery/ | free | MCP streamable-http discovery server | | POST /a2a | free | A2A JSON-RPC agent (message/send): directory search + MCP safety scan | | GET /healthz | free | Liveness | | GET /v1/models | free | Upstream model listing (read-only) | | GET /.well-known/agent-tools.json | free | Agent discovery manifest | | GET /.well-known/x402 | free | x402 v0.4 self-description (free-only) | | GET /.well-known/mcp.json | free | MCP self-description |

The MCP discovery server is also published as a standalone PyPI package: `agent-tools-mcp` (repo).

MCP discovery tools

search(intent, top_k=5, category=None, max_price_usd=None, has_mcp=None)
ask_services(intent, top_k=5, category=None, max_price_usd=None, use_llm=True)
get(slug)
list_categories()
stats()
search_mcp_servers(intent, top_k=5, chain=None, require_healthy=False)
get_mcp_server(slug)
search_a2a_agents(intent, top_k=5, x402_only=False)
search_resources(intent, protocol=None, top_k=10)
scan_mcp_safety(endpoint_url, name="", description="", tools_text="")
register(url, name=None, description=None, mcp_url=None, category=None)

search accepts natural-language intent and ranks by FTS5 + popularity + health. Each result carries a match_reason and a confidence score. ask_services uses the same retrieval-first / LLM-rerank flow as /api/v1/ask.

scan_mcp_safety vets an MCP server (by endpoint URL) for malware / prompt-injection lures before you connect: an already-indexed server returns our latest stored verdict, an unknown one is probed live, scanned, and added to the directory. The verdict comes from deterministic static rules (no code execution) over the server's advertised name + description + tool docs. Rules cover pipe-to-shell / base64-eval / PowerShell install cradles, bare-IP payload hosts, prompt-injection phrasing, and MCP tool-poisoning coercion — "always call this tool first", "before using any other tool you must…", hidden <IMPORTANT> instructions, "list all API keys / include secrets in your response", and coercion to read & forward .key/.pem/.ssh/.env files. Source-code-oriented rules (SQL/command/code injection) are intentionally not applied to natural-language descriptions to avoid false positives. Each live call also runs a frontier-LLM second opinion as an advisory dimension. It is also exposed as an A2A skill on POST /a2a. The hosted server carries the full tool set above; the stdio agent-tools-mcp PyPI package ships the core search tools only.

Deploy

Production runs behind nginx (vhost: agent-tools.cloud).

cd <install-dir>
sudo git pull
sudo systemctl restart mcpserver

systemd units live in `deploy/`: