Best 5 MCP Analytics Tools

November 8, 2026Claworc Team

Tags:mcpanalyticsintegrations

You shipped a remote MCP server. It is listed in ChatGPT and Claude, maybe in the official MCP Registry too. Then someone asks the obvious question: is anyone using it? Your web analytics can’t answer. The agent calls your tools from its own servers, answers inside the chat, and a click on a link, if there is one, arrives with no attribution.

MCP analytics tools fill that gap. They record which agents call which tools, how those calls fail, and, in some cases, what people do with the answers. If you run OpenClaw, Hermes or NanoClaw agents with Claworc, this is the traffic your agents create on other people’s servers. If you publish an MCP server yourself, it is the traffic you want to measure.

Below are the five options worth a look in November 2026, from a proxy you set up with a DNS change to SDKs you add to your server.

Key takeaways

• The main choice is where the tracking runs. An SDK runs inside your server’s code. A proxy sits in front of the server and needs no code change.
• Agent identity is harder than it looks. Most agents announce themselves, and anyone can claim to be ChatGPT. Check how a tool decides who called.
• Tool calls are only half the story. If your tools return links or widgets, you need link and widget tracking to know whether a person acted on the answer.
• Users and sessions are mostly estimates. Only ChatGPT sends a user id. Other agents don’t, so any tool that counts people across agents is estimating.

How we compared

Tool	How it is set up	Agent identification	Beyond tool calls	Hosting
MCP Tower	DNS change, no SDK	Verified signals, flags pretenders	Link tracking, UI analytics, uptime, catalog submissions	Hosted
PostHog MCP analytics	SDK in 5 languages	Client from the MCP handshake	Agent intent, missing capabilities	Hosted
TrackMCP	SDK wrapper, TypeScript and Python	Client name reported by the session	Workflow outcomes, returning clients	Not stated
AgentCat (formerly MCPcat)	Open-source SDK	Not described	Session replay, goals	Hosted or self-hosted
mcpspan	SDK in 8 languages, or a proxy	Client per call	Token cost of the tool list	Hosted or self-hosted, MIT

1. MCP Tower: MCP analytics with nothing to install

MCP Tower is a hosted middleware that sits between AI agents and your remote MCP server. You point a subdomain of your own domain at it with two DNS records, a CNAME and a TXT record for the certificate. Agents keep using your URL, every request passes through to your server, and MCP Tower records what happened. There is no SDK and no code change, and it works with any remote MCP server reachable over HTTPS that speaks Streamable HTTP.

What you get:

• Usage per agent and per tool. One page per server answers who uses it, how, from where, and whether it works. It has filters for tool, agent, country and date range.
• Agent detection that is hard to fake. MCP Tower checks HTTP message signatures, OAuth tokens, client certificates, IP ranges and networks. It records how far each label can be trusted, and it flags a request whose claimed agent its network contradicts.
• Estimated users and sessions. ChatGPT is the only agent that sends a user id. For everyone else, a statistical model links sessions into users from the signals the traffic carries. The dashboard shows those numbers as estimates, next to the counted ones, never summed with them.
• Link-click tracking. Every https:// link in a tool response is rewritten. A click is joined back to the tool call, the session and the agent that produced it, so you can see which answers sent people to your pricing or sign-up page.
• Widget tracking. For agents that render MCP widgets, loads and clicks inside the widget are recorded, even when the user never leaves the chat.
• Uptime checks. On a schedule you choose (every minute by default), MCP Tower lists your server’s tools and records the latency and any error.
• Tool changes everyone can see. Every server is rescanned daily. When a tool or argument is added, changed or removed, down to a description, Admins and Editors get an email and Inspect highlights the change.
• Inspect and edits without a deploy. Anyone on the team can read your tools in plain language and call them. Admins and Editors can rewrite tool descriptions or add prompts from the dashboard.
• Catalog submissions. You answer one questionnaire, and MCP Tower turns it into what each catalog needs: a server.json for the official MCP Registry, a plugin ZIP for ChatGPT, and form answers for Claude, Muse, Perplexity and Grok. You submit, except for the registry and the catalogs that take pull requests, where MCP Tower publishes or opens the PR when you click.
• Ask your own agent. MCP Tower’s own MCP server lets an agent, including one running in Claworc, answer “which tool fails in ChatGPT only?” or write this week’s usage report.

Best for: product and growth teams that treat their MCP server as a channel and need numbers without a data team. If you are new to the channel, MCP Tower’s free course, Growing Your MCP Server: A Crash Course, covers getting listed, measuring usage and growing it.

Trade-offs: remote servers only, since a local stdio server never passes through a proxy. Traffic goes through a hosted service.

2. PostHog MCP analytics

PostHog added MCP analytics to its product analytics suite. It “records tool calls, agent intent, reported models, failures, and requests for missing capabilities.” Each call becomes a $mcp_tool_call event next to the rest of your PostHog data.

Setup is an SDK. The getting-started guide covers TypeScript, Python, Go, Ruby and Rust: you install the package, set a project token, wrap your server and flush events on shutdown. A wizard can do it for you. The feature is labeled beta, and the Ruby and Rust integrations are experimental.

Best for: teams already on PostHog who want MCP events in the same place as their web and product events.

Trade-offs: you change and redeploy your server, and maturity varies by language.

3. TrackMCP

TrackMCP records what happens at your MCP server’s boundary: tool use, latency, errors and workflow outcomes. It groups calls into sessions and also tracks resource and prompt calls, notifications and subscriptions. The dashboard lists the assistants that use your server, such as Claude, Cursor, ChatGPT and Gemini, by the client names their sessions report, and shows active account groups and returning clients. Identifiers and prompt arguments stay redacted.

Setup is an SDK. npx @trackmcp/cli init guides it, and the integration wraps your existing server with one call, withTrackMCP(server, {...}). It works with the official TypeScript and Python MCP SDKs, with no manual event tagging.

Best for: TypeScript and Python servers that want tool and workflow analytics from a one-line wrapper.

Trade-offs: an SDK in your server, only two languages, and agent names are the ones clients report about themselves.

4. AgentCat (formerly MCPcat)

AgentCat, which was called MCPcat until recently, describes itself as analytics for Claude Connectors and ChatGPT Apps. Its focus is the session: you can replay every tool call’s request, response and errors, and see sessions grouped into the goals agents were pursuing. Errors, hallucinations and crashes are grouped by impact.

It is a hosted service with open-source SDKs, and it offers self-hosting on your own cloud for stricter data privacy. A free plan includes 500 sessions a month.

Best for: debugging what agents tried to do, one session at a time.

Trade-offs: an SDK in your server, and the site does not describe how agents are identified.

5. mcpspan

mcpspan is open-source “analytics for MCP servers,” MIT licensed, hosted or self-hosted with Docker. It shows which tools agents call, from which client, latency and failures. It also flags tools agents asked for that you don’t have, and shows how many tokens your tool list costs each session.

You can add one of its SDKs (TypeScript, Python, Go, C#, Java, Rust, Ruby, PHP) or put mcpspan-proxy in front of a server, over stdio or HTTP. Parameter values stay in your server’s process.

Best for: engineers who want an open-source tool they can run themselves, including for local stdio servers.

Trade-offs: you host and maintain it if you self-host, and its focus is on tool calls rather than what users do after the answer.

Which MCP analytics tool should you pick?

• You need to show growth numbers: users, clicks, conversions by agent. MCP Tower.
• Your product analytics already live in PostHog. PostHog MCP analytics.
• Your server is TypeScript or Python and you want a one-line wrapper. TrackMCP.
• You want to replay what agents did, session by session. AgentCat.
• You want open source, self-hosted, or analytics for a local server. mcpspan.

Conclusion

Agents are becoming a real traffic source, and an MCP server without analytics is a channel you can’t defend at the next planning meeting. Pick the option that matches who asks the questions. If it’s your growth or product lead, start with MCP Tower: organizations that sign up before January 1, 2027 get a 50% founding-customer discount for their first year.