MCP tools

Send bounded PerfChecker evidence to one MCP advice tool. Use this only if you already have an MCP server exposing a tool that can answer an advice request. To run or inspect tests, you do not need this page.

MCP is a tool protocol, not a model endpoint.

• Your server must expose a tool that can answer the request, directly or through its own assistant.

• A server with only unrelated data tools cannot generate performance advice.

• PerfChecker does not install or start that server.

Configure

Install HTTP in the dedicated controller environment. Copy examples/advisor-corpus/mcp-local.json to your project's perf/advisor.json and replace the example address and tool name. ask is a placeholder, not a standard tool.

• protocol — mcp_http.

• endpoint — local HTTP or explicitly allowed remote HTTPS.

• mcp_version — 2026-07-28, or 2025-11-25 for a legacy server.

• mcp_tool — the one advice tool; existence is checked with tools/list.

• mcp_prompt_argument — string argument receiving prompt plus evidence (usually prompt or question).

• mcp_arguments — extra arguments the tool requires; cannot replace the prompt argument.

• instructions — up to 5,000 characters; empty uses the default request.

• mcp_response — text for advice only; structured for validated references and experiment selection.

• timeout — total isolated worker time, including startup and exchanges.

For a remote server, set allow_remote = true and use HTTPS. Name the Bearer-token environment variable in api_key_env; do not store the token in the configuration. OAuth login/refresh is not implemented. A Chat Completions model uses chat_completions, not mcp_http.

using PerfChecker
config = load_advisor_config("perf/advisor.json")
result = narrate_advice(read_advice("perf/results/advice.json"); config)
display(investigation_view(result))

What is sent

• Bounded recommendation IDs, rules, observations, proposed experiments, verifications and limits.

Not sent: raw logs, source files, credentials and workspace roots.

Text replies

Text replies are limited to 16,000 characters and appear separately from deterministic findings, carrying reference_status = "unstructured_not_verified".

• Individual statements get no invented citations.

• Embedded HTML is displayed as text.

• No suggested command is executed.

Structured investigations

Keep advisorInvestigates = false for advice only. To let the assistant select experiments, set mcp_response = "structured" and enable it. The tool must return:

{
  "cards": [{"evidence_id": "an ID supplied in the request", "explanation": "Supported advice"}],
  "experiment_id": "stop"
}

• Unknown or duplicate evidence IDs and unknown experiment IDs are rejected.

• Count and duration budgets still apply; only declared experiments run.

• Text mode refuses experiment selection.

• There is no unrestricted handover, source editing, automatic baseline adoption or publication.

Compatibility and failure handling

The client supports the HTTP subset needed for one tool:

• JSON and request-scoped SSE responses;

• paginated tool discovery (32 pages max);

• legacy initialization/session release;

• modern request metadata and annotated argument headers.

Protocol revisions are explicit; there is no silent downgrade. Responses are capped at 1 MB; redirects and automatic retries are disabled.

Not implemented: stdio, deprecated HTTP+SSE endpoints, sampling, elicitation, roots, subscriptions and arbitrary tool chains.

Unsupported interactions, missing tools, errors and malformed structured replies preserve the deterministic fallback. Cancellation stops the local worker; remote interruption depends on the server.