Use MCP Tower from your AI agent

MCP Tower has its own MCP server, https://mcp.mcptower.com/. Connect it to the AI agent you already use and ask it about your MCP server in your own words: “how did last week go?”, “which tool fails in ChatGPT only?”, “why did calls drop on Tuesday?”. The same server, with the submit-mcp-to-catalog skill, lets the agent fill in a catalog’s submission form for you and stop before you submit.

The MCP server is part of paid plans: it works for MCP servers with an active subscription.

Connect the MCP server

Add https://mcp.mcptower.com/ as a remote MCP server (Streamable HTTP). The agent opens MCP Tower in your browser: sign in, read what the agent will be able to do, and press Allow. It acts as you, with your role in your organization.

• Claude (claude.ai and Claude Desktop): Settings, Connectors, Add custom connector, and paste the URL.
• ChatGPT: Settings, Apps and connectors, Create (developer mode), and paste the URL.
• Claude Code: claude mcp add --transport http mcp-tower https://mcp.mcptower.com/, then run /mcp to sign in. The plugin below adds it for you.
• Codex: add it to ~/.codex/config.toml as an MCP server with url = "https://mcp.mcptower.com/", or install the plugin below.
• Any other agent: add the URL wherever it adds remote MCP servers.

You see and disconnect every connected agent in MCP Tower under your account menu, Connected agents. Admins see and disconnect everyone’s. Disconnecting ends the agent’s access at once; it asks you to connect again.

Install the plugin

The plugin adds the MCP server and the submit-mcp-to-catalog skill in one step.

• Claude Code: /plugin install mcp-tower --marketplace mcp-tower/agent-plugin
• Codex: codex plugin marketplace add mcp-tower/agent-plugin, then codex plugin add mcp-tower.
• claude.ai: upload the skills/submit-mcp-to-catalog folder as a skill in Settings, and add the connector as above.
• Other agents that load Agent Skills: copy the submit-mcp-to-catalog folder where the agent loads skills, and add the MCP server by URL.

Ask about your MCP server

Start with “which MCP servers can you see?”: the agent calls my_mcps and uses one of its URLs from then on. Then ask anything the Measure page answers, and the next question too.

Tool	What it returns
get_summary	The eight Measure figures for a range (7 days by default), each with the period before
get_report	One Measure report: tool calls, agents, users, countries, widget and link funnels, links, latency, errors by tool, console errors
query_tool_calls	Tool calls grouped by up to three of time, tool, status, error class, agent, country, signed in, widget, and a recorded argument
uptime	Uptime, failed and slow checks, downtime, check times, the timeline and each outage

In Claude and ChatGPT the results also show as charts, with buttons for 24 hours, 7 days and 30 days and a link to the same view in MCP Tower.

How to read the numbers:

• Users are estimates, shown as a range. Don’t add them up across days.
• The newest minutes are approximate: a dip in the last bucket is often not real yet.
• A notice that tracking is off means the data was not recorded, not that it is zero.

What leaves MCP Tower

Results go to your agent and its model provider. They are overall numbers only: counts, rates, percentiles and shares over a group. No result names a user, person or session, and none contains an IP address, a user agent, an exact link URL or a single event. Argument values appear only as group labels, and only for the arguments you chose to record in Tracking. Your own privacy terms with your users apply to what you share with your agent.

Submit to a catalog

Ask “submit my MCP server to Claude”, or run /submit-mcp-to-catalog claude. The skill:

1. lists your servers and the catalogs, with each one’s status;
2. checks that your server is ready, and asks you for any missing answer, which it saves to your Distribute answers in MCP Tower;
3. shows you every value it will enter, and continues when you say so;
4. opens the catalog’s form in your browser, where you sign in, and fills it in;
5. stops at every checkbox you must tick and at the final Submit button: you press them, or tell it to after reading them;
6. records the submission in MCP Tower.

The MCP registry is published by MCP Tower itself, from the Publish to registry button in Distribute: the skill shows you the server.json preview and the link to that button.

Viewers and Editors

Your role	Your agent can
Viewer	Analyze every server, see the catalogs, their checks and submission data
Editor or Admin	Also save answers, read a catalog’s test-account password into its form, and record submissions

Test-account passwords reach your agent and its transcript, so use review accounts with limited access. Each password the agent reads is recorded in your audit log.