[Feature Request] Native Model Context Protocol (MCP) server support inside standalone Pyroscope binary

[mahigadamsetty]

Is your feature request related to a problem? Please describe.

Yes. With the explosive rise of AI-native development workflows (Cursor, Claude Desktop, Windsurf, Cline, VS Code Copilot), engineers increasingly rely on AI agents to diagnose and fix performance regressions.

Currently, the only official way to connect an LLM agent to continuous profiling data is via the grafana/mcp-grafana repository. This requires a full Grafana stack or a Grafana Cloud orchestration layer to route queries. For organizations running a lean, standalone open-source pyroscope container (without the UI footprint of Grafana), this creates an unnecessary, heavy middleware dependency and introduces a complex configuration bottleneck.

Describe the solution you'd like

I would like to request native, lightweight Model Context Protocol (MCP) server capabilities compiled directly into the core pyroscope binary.

Running pyroscope --mcp-server or configuring a minimalist standalone flag should expose basic MCP tools:

1. list_profiles - Discover profile types (CPU, memory, goroutines, mutex, etc.)
2. get_profile - Expose the profile call-tree in a text/collapsed/DOT format directly to local AI clients.

Describe alternatives you've considered

1. Using the official mcp-grafana server, which forces the usage of Grafana enterprise/cloud infrastructure tokens and complex data source UID mapping.
2. Manually downloading pprof/collapsed text profiles via the raw HTTP endpoint and copy-pasting them into LLMs, which is slow and breaks the automated context-loop of modern AI agents.
3. Writing local custom Python FastMCP server shims to wrap around the raw HTTP API.

Additional context