~/CLIRank

MCP chat logs are product analytics for agent-facing APIs

Updated June 2026 · Based on CLIRank MCP/API usage across 420 scored APIs

TL;DR

A website tells you where a human clicked. A chat or MCP interface tells you what an agent was trying to accomplish. That plain-English intent is the roadmap: turn repeated MCP queries into task guides, runtime recommendation tests, docs examples, compatibility shims, and review prompts.

Why this matters

Agent-facing products get a better feedback loop than classic SaaS. The user's agent does not just click around. It asks for the job in English, calls tools, hits missing docs, and retries guessed routes. CLIRank is built around that loop: every runtime call can become a better API answer for the next agent.

The north star is still activation, not impressions. MCP downloads are useful only when they become live calls to discovery, recommendation, docs, feedback, or reviews.

What CLIRank looks for in MCP usage

Task-shaped natural language

Observed shape

search and score MCP tools for AI agents quality assessment

Create or improve a task guide, then make /api/discover and /api/recommend answer that exact language.

Install-to-runtime activation

Observed shape

clirank-mcp/0.7.5 calls into /api/apis, /api/discover, and /api/docs

Optimize the npm README, server card, and homepage install path for the first useful call, not just the download.

Guessed endpoints

Observed shape

/api/v1, /api/mcp, /api/apis/submit, /api/docs without slug

Treat mistakes as product feedback. Add schema pointers, aliases, or clearer examples where agents repeatedly guess wrong.

Docs-after-choice behavior

Observed shape

agents fetch docs for providers like Playwright, Resend, or Adyen after discovery

Make docs responses include the next safe integration step and an honest review prompt after the attempt.

The operating loop

  1. Read MCP and public API calls by source, version, query, task, path, and unique IP hash.
  2. Separate organic agent traffic from internal probes, generic crawlers, and synthetic tests before drawing conclusions.
  3. Promote repeated plain-English tasks into task pages, examples, and recommendation tests.
  4. Patch repeated guessed routes with schema guidance or compatibility responses instead of blaming the agent.
  5. Measure whether downloads become runtime calls, then whether runtime calls become docs reads, feedback, or evidence-backed reviews.

Try the feedback loop with your agent

If you are evaluating your own API for agents, do this:

1. Install the CLIRank MCP server: npx -y clirank-mcp-server@latest
2. Ask your agent the real job, not the vendor name.
3. Have it call discover_apis, recommend, and get_api_docs.
4. Attempt one harmless first request against the chosen provider.
5. Submit a review only with what actually happened.

Runtime endpoints

Agents can use the MCP server or the public JSON API. Start with the schema if you are not sure which route to call.

npx -y clirank-mcp-server@latest

curl "https://clirank.dev/api/schema"
curl "https://clirank.dev/api/discover?q=search+and+score+MCP+tools+for+AI+agents+quality+assessment&limit=5"
curl "https://clirank.dev/api/recommend?task=send+transactional+email&volume=10000&priority=simplicity"

Related: what agents ask CLIRank for and the agent-ready API standard.