MCP

Connect Mida to your AI agent

Mida runs a hosted MCP server. Point your agent at it, sign in with your Mida account, and it can build tests, read results, and roll out winners without you switching tabs.

The server

One remote MCP server, authorized in your browser with your Mida account. No API keys to paste, no separate seat to buy.

Server URL https://mcp.mida.so/mcp

Using Cursor or VS Code? Install it in one click:

In a terminal agent, paste one command:

Claude Code
claude mcp add --transport http mida https://mcp.mida.so/mcp
Codex
codex mcp add mida --url https://mcp.mida.so/mcp codex mcp login mida

Run both. Codex only opens the Mida sign-in on the second command, and stays disconnected until it has run.

Most other clients take a small block of JSON. If yours does, this is it:

{ "mcpServers": { "mida": { "type": "http", "url": "https://mcp.mida.so/mcp" } } }

Transport is streamable HTTP. Auth is standard MCP OAuth with dynamic client registration, so your client registers itself and opens a browser the first time you connect. Nothing is stored on your machine.

Want it read only? When you sign in, choose Read only. Your agent can then see experiments, results and settings but cannot create, change or delete anything. Useful for stakeholders and client accounts.

Set up your client

Step by step instructions for the clients people use most. Any other MCP client works too, using the server URL above.

Run Mida inside your workflows

Automation tools can call Mida too. Pull results on a schedule, post a weekly summary to Slack, or let an agent in Teams answer questions about your tests. Each one signs in with your Mida account, with no API key to copy.

Before you start

Three things worth knowing, in the order they bite.

  1. 1. The tag has to be on your site

    Mida assigns visitors to variants from a small script on your pages. Nothing runs until it is installed. Ask your agent for the snippet with get_install_snippet, or grab it from your Mida dashboard. If a live test reports no visitors at all, check the tag before you check the targeting.

  2. 2. Pick the right project first

    One Mida login can own several organizations and projects, and the session opens on a default project that is often not the one you mean. Your agent checks which project it is on when the chat starts and tells you. If it is the wrong one, ask it to switch, or run list_projects and select_project yourself. A test in another project comes back as 404: Experiment not found, which reads like a deleted test when it is really a scoping miss.

  3. 3. Your agent can write, not just read

    67 tools and 11 MCP prompts cover the whole loop: create experiments and goals, generate a preview URL, launch, read results with Mida's own statistics, conclude a test with a written conclusion, and serve the winner. New tests are created as drafts, and anything visitors would see waits for your yes in the chat. If your role needs approval to publish, your agent can send the change to an Owner or Admin instead.

Give your agent something to test

Free up to 100,000 tracked users a month. No credit card, no sales call.