davi17g/parcaprof-mcp

By davi17g

•Updated 4 months ago

MCP server bridging Parca continuous-profiling to LLM clients.

Image

Developer tools

Monitoring & observability

426

Overview Tags

davi17g/parcaprof-mcp repository overview

⁠parcaprof-mcp

An MCP⁠ server that exposes a Parca⁠ continuous-profiling backend to MCP clients (Claude Desktop, Claude Code, MCP Inspector, …) so an LLM can explore profile types, label sets, time-range series, and per-function hotspots (top / flamegraph) over a Parca server's gRPC API.

Built with the official modelcontextprotocol/go-sdk⁠.

Source: https://github.com/davi17g/parcaprof-mcp⁠

⁠Image

Registry: docker.io/davi17g/parcaprof-mcp
Tags: vX.Y.Z per git tag, plus :latest on the most recent tag
Architectures: linux/amd64, linux/arm64
Base: alpine, runs as non-root user parcauser (UID 65532)
Exposes HTTP on :8080 (Streamable HTTP at /mcp, SSE at /sse, liveness at /healthz)

⁠Quick start

Run against a local Parca on the host (plaintext gRPC):

docker run --rm -p 8080:8080 davi17g/parcaprof-mcp:latest \
  --parca-address=host.docker.internal:7070 \
  --parca-insecure

Then point an MCP client at http://localhost:8080/mcp.

⁠Transports

MCP transport	Endpoint	Notes
Streamable HTTP	`/mcp`	Current MCP spec; preferred for remote clients
SSE	`/sse`	Legacy spec; kept for older clients
stdio	—	Use the binary directly (not this image) for Claude Desktop

gRPC is used only to talk to Parca. There is no MCP-over-gRPC transport.

⁠Tools

Tool	Parca RPC
`parca_profile_types`	`ProfileTypes`
`parca_labels`	`Labels`
`parca_label_values`	`Values`
`parca_series`	`Series`
`parca_query_range`	`QueryRange`
`parca_query_single`	`Query` (compact summary)
`parca_top`	`Query` with `REPORT_TYPE_TOP`, sorted top-N rows
`parca_flamegraph`	`Query` with `REPORT_TYPE_FLAMEGRAPH_TABLE`, flattened & aggregated

Times accept RFC3339 or relative now, now-15m, now-1h30m, now+5m. The default window is the last 15 minutes.

⁠Flags

Flag	Default	Description
`--parca-address`	`localhost:7070`	Parca gRPC `host:port`
`--parca-insecure`	`false`	Use plaintext (no TLS) to Parca
`--transport`	`http`	`http` (Streamable HTTP + SSE) or `stdio`
`--http-addr`	`:8080`	HTTP listen address (when `--transport=http`)
`--version`	—	Print version, commit, and build time, then exit

⁠Environment

PARCA_BEARER_TOKEN — sent as authorization: Bearer … on every gRPC call to Parca.

⁠TLS to Parca

By default the server dials Parca over TLS. Use --parca-insecure only for local/dev Parca instances. For a managed Parca behind TLS + bearer auth:

docker run --rm -p 8080:8080 \
  -e PARCA_BEARER_TOKEN=$PARCA_TOKEN \
  davi17g/parcaprof-mcp:latest \
  --parca-address=parca.example.com:443

⁠Health check

GET /healthz returns 200 OK for liveness/readiness probes.

⁠Verify end-to-end

Start Parca: docker run --rm -p 7070:7070 ghcr.io/parca-dev/parca:v0.28.0
Start this server (see Quick start above)
npx @modelcontextprotocol/inspector → connect to http://localhost:8080/mcp → list tools → call parca_profile_types

⁠License

Apache-2.0⁠

Tag summary

Recent tags

Content type

Image

Digest

sha256:9564c4aff…

Size

12.4 MB

Last updated

4 months ago

docker pull davi17g/parcaprof-mcp