maxoperf

⚡ MaxoPerf MCP Server

Drive the entire MaxoPerf platform from any AI agent — with your API key.

A remote, hosted Model Context Protocol server that turns Claude, Cursor, Codex, VS Code Copilot, and ChatGPT into a full performance-testing operator. Create tests, launch load runs on managed cloud runners, read results, and get root-cause analysis — all in natural language, all authenticated with your own key.

https://app.maxoperf.com/mcp

---

Configuration

It's a remote, hosted server — nothing to install or run. Add this to your MCP client config and set MAXOPERF_API_KEY (create one in the console → Settings → API keys):

{
  "mcpServers": {
    "maxoperf": {
      "type": "http",
      "url": "https://app.maxoperf.com/mcp",
      "headers": {
        "Authorization": "Bearer ${MAXOPERF_API_KEY}"
      }
    }
  }
}

Your key is validated on every call by the real platform (auth, tenancy, OpenFGA, audit) — the server stores nothing and adds no new trust boundary. Revoke the key, access dies instantly. Send the header X-MaxoPerf-MCP-Mode: read-only for a look-but-don't-touch session (write tools hidden).

Install (pick your client)

Claude Code — the plugin bundles the server and the agent skill:

/plugin marketplace add MaxoPerf/mcp
/plugin install maxoperf

Or add just the connector:

claude mcp add --transport http maxoperf https://app.maxoperf.com/mcp \
  --header "Authorization: Bearer ${MAXOPERF_API_KEY}"

Cursor · VS Code / Copilot · Codex · ChatGPT · Claude Desktop — one-click deeplinks and copy-paste config in packaging/. All use the same mcpServers block above.

---

Tools

29 curated tools. Reads default to response_format: "concise" (pass "detailed" for the full payload); write tools require a non-read-only session, and cancel_run is hidden in read-only mode.

Context & tenancy

• whoami — Resolve the account + default workspace behind your API key
• list_workspaces — List workspaces visible to the account
• set_active_workspace — Set the active workspace for the session
• list_projects — List projects (with edit/delete permissions)
• create_project — Create a project

Tests

• list_tests — List tests (filter by project / workspace / type)
• get_test — Get one test + its validation summary
• create_test — Create a test shell (choose the engine/executor)
• get_test_overview — Run-history overview for a test

Test files

• upload_test_file — Upload a script/data file in one call (real 3-step presigned flow)
• list_test_files — List a test's files + upload state
• download_test_file — Get a short-lived download URL for a file

Runs

• start_run — Launch a load/browser run on managed cloud runners (idempotent)
• get_run_status — Poll lifecycle status (queued → running → passed/failed/cancelled)
• list_runs — Paginated run history with filters
• cancel_run — Cancel a run (destructive; hidden in read-only)
• rerun_run — Re-run from a snapshot or the current test
• add_runners — Scale a live run up at existing locations

Results

• get_run_results — KPI overview: throughput, latency percentiles, error rate
• query_run_metrics — Time-series metrics (latency / throughput / errors / load / health)
• get_run_errors — Grouped error rows (message / count / code)

Diagnostics — root-cause & anomaly detection

• get_run_summary — Executive summary + which failure criteria tripped
• get_run_error_bodies — Sampled error request/response bodies + status codes
• get_run_logs — Error-level engine/system log lines
• get_runner_health — Runner CPU/mem trend + targetVus vs peakAchievedVus (vuShortfallPct)
• detect_run_anomalies — Deterministic robust-outlier scan (median/MAD); terminal-gated, low false-positive

Escape hatch & discovery

• call_platform_api — Reach any public /v1/* endpoint (secrets, environments, schedules, BYOC); admin/internal deny-listed, SSRF-safe
• get_openapi — The public OpenAPI document
• search_endpoints — Keyword search over the API to find the right endpoint

Prompts

Text recipes that encode the correct tool sequence — great for chat-only clients that can't read a repo:

• run-baseline-load-test — Start a baseline run and watch it to completion
• diagnose-latency-regression — Compare p95 across two runs
• summarize-run — Plain-language summary of one run
• plan-and-build-test — Turn a goal into project → test → upload → run
• choose-executor — Recommend an engine (k6 / JMeter / Playwright / Selenium)
• scan-endpoints-for-hotspots — Rank likely hotspots from an OpenAPI spec or pasted list
• setup-secrets-and-envs — Wire workspace secrets + multi-env before a run
• diagnose-run-failure — Ranked root cause for a failed run
• explain-run-anomalies — Explain each detected outlier

Resources

• maxoperf://openapi — The public OpenAPI spec
• maxoperf://run/{id} — A run report summary

---

Try it

• "Load test https://api.example.com/checkout with 500 users for 5 minutes and fail it if p95 goes over 800ms."
• "Scan my repo for the endpoints most worth load-testing, then build and run a test for the riskiest one."
• "Why did run run-0000000001 fail? Check the errors, the logs, and whether the runners actually reached the target load."

Pair it with the brain

The MCP server is the hands. The bundled MaxoPerf agent skill (npx @maxoperf/agent-skill install, included in the Claude plugin, or in agent-skill/ here) is the brain — it reads your code, finds the hotspots, builds and runs the test, and diagnoses why it broke, driving these tools automatically.

What's in this repo

Path	What
.claude-plugin/	Claude Code plugin (bundles the MCP connector + the skill)
agent-skill/	A copy of the maxoperf agent skill (canonical home: MaxoPerf/agent-skill)
packaging/	MCP Registry server.json, .mcpb, VS Code / Cursor deeplinks, Codex / ChatGPT setup
LAUNCHGUIDE.md	MCP directory listing metadata

---

Get started → · Agent skill → · Docs →

Read-only mirror of the MaxoPerf monorepo. File issues at maxoperf.com.