MCP Server

The Monito MCP server lets an AI tool run Monito directly: it finds or writes a test scenario, runs it in a real browser, reads the bug report, fixes the code and runs it again. Everything the CLI does for agents is available as typed MCP tools.

Two ways to connect

Use the hosted server unless you want the CI setup tools or your client can only run local servers. Both expose the same tools and act as your Monito account.

Quick start

In Claude Code:

claude mcp add --transport http monito https://www.monito.dev/api/mcp

Open Claude Code, run /mcp and choose monito to sign in. Your browser opens, you approve the connection, and the tools are ready. See Connect a client for Claude, Cursor, VS Code, ChatGPT and the local server.

Then ask:

Check that the checkout on https://staging.example.com still works after my change, and fix anything Monito finds.

What a session looks like

1. The agent calls list_scenarios to find a scenario for the page, or create_scenario with a plain-language prompt.
2. run_scenario runs it in a real browser and waits for the verdict. It returns the run, the number of bugs, the credits charged and a link to the run report.
3. On a failed run, get_run returns the bug reports with reproduction steps, failed network requests, console errors and screenshot names. get_screenshot shows a screenshot to agents that can see images.
4. The agent fixes the code and calls run_scenario again.

When a run errors or the scenario itself looks wrong, diagnose_run asks Monito's assistant whether the app or the instructions are at fault, and lint_scenario and improve_scenario tighten the instructions before the next run.

The server also offers a verify-change prompt that walks an agent through this loop. In Claude Code it appears as /mcp__monito__verify-change.

Runs, waits and credits

• Every run spends credits, exactly as it does from the dashboard or CLI. get_credits shows the balance.
• Run tools wait up to four minutes for a verdict. If the run takes longer, the tool returns the run ID with done: false; the agent calls wait_for_run to keep waiting. The run is never started twice.
• Pass a requestId (any unique string, such as a UUID) to run_scenario or run_project. If your client retries the call, Monito returns the first launch instead of charging again.
• A failed run means the agent found a bug. An error run is not a verdict about your app; get_run_logs shows what went wrong.

Errors

Tool errors carry the same stable codes as the CLI: auth, not_found, validation, insufficient_credits, plan_required, timeout, network, server and conflict. Credit and plan errors include a link to manage your plan. Tools that delete data require confirm: true.