MCP 服务器

chiark-mcp

Find the most reliable AI agent for any task. Search 2,000+ agents across A2A and MCP with quality filters — min uptime, max latency, score thresholds. Check if an agent is alive before routing to it. Like Artificial Analysis, but for agent services.

README

Chiark MCP Server

MCP server for AI agent discovery and quality scoring. Find reliable agents across A2A and MCP ecosystems with quality constraints.

Powered by chiark.ai — the cross-protocol quality index for AI agent services, tracking 2,000+ agents from 9 registries with three-tier operational scoring.

Quick Start

Use the hosted endpoint (recommended)

Add to your MCP client config (Claude Code, Cursor, etc.):

{
  "mcpServers": {
    "chiark": {
      "url": "https://chiark.ai/mcp/"
    }
  }
}

Install locally

pip install chiark-mcp

Add to your MCP client config:

{
  "mcpServers": {
    "chiark": {
      "command": "chiark-mcp"
    }
  }
}

Or run directly:

python -m chiark_mcp

Tools

find_agent

Search for AI agents by task description with quality constraints.

find_agent(
  task_description="web scraping",
  min_uptime=0.95,
  max_latency_ms=500,
  protocol="mcp",
  max_results=5
)

Returns ranked agents with scores, uptime, latency, endpoint URLs.

check_agent_status

Check if an agent is alive right now.

check_agent_status(agent_id="uuid-from-find-results")

Returns: is_alive, HTTP status, response time, TLS validity, last probe timestamp.

get_agent_score

Get full quality score breakdown.

get_agent_score(agent_id="uuid")

Returns: availability (0-30), conformance (0-30), performance (0-40), uptime, latency, trend, rank.

report_outcome

Report whether a routed agent succeeded or failed. Improves future recommendations.

report_outcome(agent_id="uuid", success=true, task_category="translation")

get_ecosystem_stats

Get ecosystem overview: total agents, online count, average scores, top categories.

get_ecosystem_stats()

How It Works

Chiark crawls 9 public agent registries every 24 hours and probes every discovered agent every 30 minutes across three tiers:

Availability — Is it alive? HTTP status, response time, TLS
Conformance — Does it follow its declared protocol correctly?
Performance — How fast does it respond? Task completion rate

Agents are scored 0-100 (or 0-45 for auth-gated agents that can't be fully tested).

Constraint Filters

Parameter	Description	Example
`min_score`	Minimum operational score (0-100)	`50`
`min_uptime`	Minimum 30-day uptime (0-1)	`0.99`
`max_latency_ms`	Maximum P95 latency	`500`
`auth_required`	Filter by auth requirement	`false`
`payment_enabled`	Filter by x402 payment	`true`
`protocol`	`a2a` or `mcp`	`mcp`
`category`	Skill category	`Developer Tools`