MCP server
AgentStats runs an MCP server at https://agentstats.io/mcp (Streamable HTTP). Connect it to Claude Code, Cursor or any MCP client so your coding agent can install the SDK for you, or answer questions like "which tools failed most this week?". Every tool only reads, except update_recommendation, which records that a recommendation was done, and run_evals, which starts eval runs.
Connect
claude.ai, ChatGPT and other connectors: add https://agentstats.io/mcp as a custom connector. When a tool needs your data, the connector asks you to sign in to AgentStats and allow read access for your team (OAuth, as the MCP spec describes). You can revoke it any time by signing out of the connector.
Claude Code, Cursor and clients that take headers: create a team API key under Account settings → API keys (al_sk_…), then add the server to your client:
claude mcp add --transport http agentstats https://agentstats.io/mcp --header "Authorization: Bearer al_sk_…"
{
"mcpServers": {
"agentstats": {
"url": "https://agentstats.io/mcp",
"headers": { "Authorization": "Bearer al_sk_…" }
}
}
}
Signing in, or a key, is optional. Without either you only get the setup tools.
Tools
| Tool | Needs a key | What it does |
|---|---|---|
get_setup_instructions | No | Install steps for a given framework. With a key, they include your real public keys |
score_webmcp_tools | No | Score tool definitions 0–100 for agent-readiness, with concrete issues |
list_sites | Yes | Your team's websites, with public keys and tool counts |
get_overview | Yes | Calls, sessions, success rate and latency with trends, top tools and recent errors |
get_tool_detail | Yes | One tool in depth: latency, error groups, recent calls, schema history |
get_errors | Yes | Failed calls grouped by tool and error message |
get_schema | Yes | The live tool surface, its readiness score and recent schema changes |
run_funnel | Yes | Drop-off across a sequence of 2–8 tools |
list_sessions | Yes | Recent agent sessions and the tools each one called |
get_visitors | Yes | The people on the website: visitors, sources, UTM tags, pages, countries, devices and goals, with filters |
get_traffic | Yes | Human visitors vs. agent sessions by day, plus AI-assistant referrals |
get_seo | Yes | Google and Bing search: clicks, top queries and pages, indexing suggestions and page speed |
get_ai_bots | Yes | AI crawler and agent visits by product and category |
get_ai_visibility | Yes | What ChatGPT, Claude, Gemini, Perplexity and Google AI Mode say about your product, from the latest AI visibility run |
get_recommendations | Yes | Everything worth doing in the project, ranked, with each one's status (open, planned) |
update_recommendation | Yes | Marks a recommendation planned, done or not for us (or back to open), after you've acted on it |
get_agent_browsers | Yes | Agents browsing your pages in a real browser (Claude in Chrome, automation) |
get_docs_traffic | Yes | How agents read your docs: Markdown served, most-read pages, pages that don't exist |
get_mcp_sign_in | Yes | OAuth sign-ins on your MCP servers, by client app and scope |
get_feedback | Yes | What agents and users sent through an MCP server's send_feedback tool |
list_evals | Yes | An MCP server's evals with each agent's latest verdict |
run_evals | Yes | Starts an MCP server's evals against the live server |
get_eval_run | Yes | One eval run: verdict, score, problems with fixes, every tool call |
llms.txt
For agents that don't speak MCP, /llms.txt has the same setup instructions as plain Markdown. Paste the link into any chat and ask it to install AgentStats.