MCP server — connect any AI agent to MaxoPerf
MaxoPerf runs a Model Context Protocol (MCP) server. Through it, an AI agent in your IDE, in a chat client, or MaxoPerf’s own Loadrigo can act on your account the way you would by hand. It can build a test, start a run, poll it, and read back KPIs and errors, and you ask for all of it in plain language.
The endpoint
Section titled “The endpoint”https://app.maxoperf.com/mcpThe server uses MCP’s Streamable HTTP transport. Any client that supports that transport can connect, and you don’t need a MaxoPerf-specific SDK.
The console shows the same endpoint under Settings → Connect AI tools, already resolved for the deployment you’re signed in to:
Authentication
Section titled “Authentication”Every request needs two things:
| Header | Value |
|---|---|
Authorization | Bearer mpak_<keyId>_<secret>, a MaxoPerf API key |
X-Account-Id | Your account id, if your API key’s principal has access to more than one account |
An API key grants full-account access. An agent that holds the key can do anything you can do signed in to the console. Treat the key as a credential. Name each key so you can tell them apart (claude-code-agent, ci-nightly), and revoke a key immediately if it leaks.
Read-only mode
Section titled “Read-only mode”A client can limit itself, or be limited, to read-only tool calls by sending X-MCP-Read-Only: true on every request. In read-only mode:
- Read tools (
list_runs,get_test,get_run_results, …) work normally. - Write tools (
start_run,create_test,deploy_virtual_service, …) return aREAD_ONLY_MODEerror instead of running. The agent still sees that the tool exists, but the session can’t act. cancel_runis destructive and can’t be undone, so a read-only session’s tool catalog doesn’t list it at all. It doesn’t just return an error.call_platform_api(the escape hatch) accepts onlyGETrequests.
Use read-only mode when you want an agent to investigate and report with no risk of changing anything.
Response sizing
Section titled “Response sizing”Read tools return a concise response shape by default, with small field sets an agent can read easily. They also truncate large time series to stay well inside typical context budgets. Pass response_format: "detailed" on a call when you need the full payload.
What’s available
Section titled “What’s available”The server exposes three kinds of things over MCP. Tools make up most of the surface; Tools lists the full catalog. Prompts are multi-step recipes, such as “run a baseline load test and report the result”. Two read-only resources, maxoperf://openapi and maxoperf://run/{id}, serve clients that prefer resources to tool calls.