What the server is
Supervisor Autotune is an MCP server. MCP, the Model Context Protocol, is the way AI agents such as Claude, ChatGPT and Claude Code use outside services. Once your agent is connected, it can start load tests, monitors and accessibility checks on your site and read the results, and you can ask for all of it in plain language.
- The address is
https://api.supervisor.com/mcp. - The server gives your agent 40 commands, which MCP calls tools, in six groups, and two prompts.
- These docs describe version 1.11.0 of the server.
- Everything an agent reads or changes belongs to the organization you signed in to. It cannot see other organizations.
Connect
There are two ways in, depending on your agent.
- Sign in. Claude, ChatGPT and Claude Code connect by signing in to Supervisor. You add the address as a connector, sign in and choose your organization.
- Access token. Other MCP clients send an organization access token with the scope
autotunein theAuthorization: Bearerheader. You create the token on the Agents page in the app.
The step-by-step instructions, with the address and the Claude Code command to copy, are on the Autotune page. You can disconnect an agent at any time in the app.
What a command answers
Every command answers with structured data that the agent reads and then explains to you. Some answers also carry images: charts of a load test, or screenshots of a failing element.
- Many answers end with a
nextorhintline that tells the agent what to call next, for example to ask for the status again in 15 seconds. - When something cannot be done, the answer is one plain sentence that says why and what to do, for example
example.com is not verified for this organization. Call verify_site first. - Commands that only read are marked read-only. Commands that send traffic to your site, cost credits or delete something are marked so that an MCP client can ask you first. Whether it asks is up to the client, so the prompts below also tell the agent to ask.
The autotune prompt
autotune is a step-by-step method for load testing and tuning a site. It takes two optional values: url, the site to tune, and target_users, the goal, for example 100. It tells the agent to:
- Check with
list_sitesthat the domain is verified, and if it is not, callverify_siteand tell you exactly what to add. - Find your server logs and write down the current limits, such as database connections and worker counts, before any test.
- Ask you before starting load. Then run a baseline with
start_load_test, follow it withget_run_statusand read it withget_run_result. - Read the server logs for the time of the run and call
get_slow_requeststo see which requests were slow. - Propose exactly one change, apply it only after you approve, and record it with
report_change. - Run again with the same number of users and call
compare_runs. Keep the change if it helped and revert it if it did not. - Once the site reaches the target cleanly, run once more with more users to find its real capacity.
- Stop when the goal is reached, or when what is left needs a bigger server or code changes, and say so.
- Write a journey when the important flows are not plain pages, such as search, add to cart or signing in.
- Offer to keep watching with a scheduled load test or a monitor.
The prompt ends with two rules: never change two things between runs, and never load test a site you do not own.
The accessibility prompt
accessibility is a method for finding and fixing a site’s accessibility issues and keeping its accessibility statement honest. It takes one optional value, url. It tells the agent to:
- Start from what is already known:
list_accessibility_issuesandget_accessibility_statement. It runs a new check only when there are no issues yet or the last check is old. - Check a journey when the pages that matter are behind a cart, a form or a sign-in, because a crawl does not sign in or fill forms.
- Find each issue’s element in your code and fix the cause in the component or stylesheet, not on one page.
- Decide the issues marked “needs review” with
update_accessibility_issue, and ask you when it is unsure. - Fix the most serious and most widely shared problems first, show you each change and apply it only after you approve. It must not hide problems from the checker.
- Verify the fixes with a small check of the affected pages.
- Test with you what automated checks cannot: keyboard use, 200% zoom, a 320 px wide window, form errors and captions. It records what fails with
add_accessibility_issue. - Never publish the statement without asking you, and never claim that the site is fully compliant from automated results.
How to use a prompt
- In an MCP client that lists prompts, pick
autotuneoraccessibilityfrom the list. In Claude Code they appear as slash commands, for example/mcp__supervisor__autotune. - Or just say what you want. The server tells every connected agent that these two methods exist, so a plain request is enough.
“Use Supervisor to find out how many users example.com can handle. The goal is 200.”
“Check the accessibility of example.com with Supervisor and help me fix what it finds.”