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

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

Currently deployed on latex-tools behind nginx (vhost: agent-tools.cloud).

cd /opt/mcpserver
sudo git pull
sudo systemctl restart mcpserver

systemd units live in deploy/:

• mcpserver.service — the ASGI app (uvicorn, 127.0.0.1:9100)
• agent-tools-crawl.{service,timer} — 6h directory crawler
• agent-tools-health.{service,timer} — endpoint health checks

Config

See .env.example. Required:

• UPSTREAM_BASE_URL / UPSTREAM_API_KEY — used by /v1/models
• AGENT_TOOLS_ASK_BASE_URL / AGENT_TOOLS_ASK_API_KEY / AGENT_TOOLS_ASK_MODEL — LLM backend for /api/v1/ask
• AGENT_TOOLS_SAFETY_BASE_URL / AGENT_TOOLS_SAFETY_API_KEY / AGENT_TOOLS_SAFETY_MODEL — optional override for the scan_mcp_safety advisory LLM (falls back to the AGENT_TOOLS_ASK_* backend)
• AGENT_TOOLS_DB_PATH — SQLite path for the directory
• AGENT_TOOLS_ASK_RATE_LIMIT_PER_MINUTE / AGENT_TOOLS_ASK_RATE_LIMIT_PER_DAY — public ask abuse limits
• METRICS_BEARER_TOKEN — optional remote access token for /metrics; without it metrics are local-only

Smoke test

# Liveness
curl https://agent-tools.cloud/healthz

# Directory stats
curl https://agent-tools.cloud/api/v1/stats

# Intent-level service recommendation
curl -s -X POST https://agent-tools.cloud/api/v1/ask \
  -H 'content-type: application/json' \
  -d '{"query":"find a weather API that accepts x402","limit":2}'

# Service card
curl https://agent-tools.cloud/api/v1/services/weather-hugen-tokyo-scan

# MCP discovery handshake
curl -s -X POST https://agent-tools.cloud/mcp-discovery/ \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-06-18","capabilities":{},
                 "clientInfo":{"name":"smoke","version":"1"}}}'

# Retired paid path -> 404
curl -s -o /dev/null -w '%{http_code}\n' -X POST \
  https://agent-tools.cloud/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"Qwen/Qwen3.6-35B-A3B","messages":[{"role":"user","content":"hi"}]}'
# => 404

License

The agent-tools.cloud server and directory code in this repository is licensed under the PolyForm Noncommercial License 1.0.0 — you may use, modify, and share it for any noncommercial purpose; commercial use is not permitted.

The standalone MCP client package in agent-tools-mcp/ (published to PyPI as agent-tools-mcp) is licensed separately under Apache-2.0, so any agent — including commercial ones — can install and call the hosted service freely.

Using the hosted service at agent-tools.cloud is governed by its terms of service, not this code license.