Parca MCP server: Polar Signals Cloud only

Why the vendor server needs Polar Signals Cloud, what its eight tools do, what self-hosted Parca can use instead, and the context budget Polar Signals publishes.

Checked 3 October 2026 · from Polar Signals release posts, the live endpoint and the GitHub API

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

What is the Parca MCP server?

Continuous profiling samples what your processes are actually doing, all the time, down to the line number. The output is a flame graph: which function called which, and where the time went.

A flame graph is a difficult thing to hand a model. Polar Signals said so directly in the 30 June 2026 post. A flat table of several hundred rows forces the model to reconstruct the call structure in its head. A full flame graph dumped into context has the same problem, only bigger.

Polar Signals designed for navigation. The agent moves through the graph instead of receiving all of it. The sandwich tool returns two trees rooted at one function, its callers going up and its callees going down. The flame graph tool takes a focus_path so the model zooms into one branch and the siblings are trimmed away.

Profiling is a narrow slice, and this connector fills it more deeply than anything else here.

For an agent, that means Parca answers one question extremely well: which function is expensive, and who calls it. It cannot tell you that a request failed, only what the CPU was doing while it did.

What can an AI agent do with Parca?

What the tool definitions cost in context

A client that loads every tool definition at the start of a session spends this much context before the first question. Claude Code defers definitions by default through tool search, and loads a tool's definition when the agent needs it. The method below gives the version and date measured.

A context window is the amount of text a model can hold at once. Eight tools cost 780 to 880 tokens on Claude. The descriptions are short and uniform, so the two tokenizer generations barely differ. Only Split.io has a narrower range, at 912 to 983.

The vendor tools exist only on Polar Signals Cloud, so this figure applies only to the hosted route. The community server for self-hosted Parca has its own eight tools, not measured here.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Measured on or before 25 August 2026.

Counted from the 8 tool entries in the vendor's own documentation. That is the only public inventory.

The server could not be run for a live tools/list. So no input schemas are available. These figures cover each tool's name and description only. They are a floor, not the figure.

The Claude columns are measured through the API's count_tokens endpoint, with an empty schema per tool. So they cover the same content as the GPT columns. The range there is two tokenizer generations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a tenth.

Parca MCP server limits: where answers come back incomplete

Parca's main boundary rules out most people who run the product, before its other limits matter.

How to configure Parca MCP for agents

1. 1

   Confirm you are on Polar Signals Cloud

   The vendor server reads only Polar Signals Cloud projects. Self-hosted Parca needs the community davi17g/parcaprof-mcp, which queries Parca's gRPC API instead.

2. 2

   Let the skill load, then discover before you query

   The server serves a continuous-profiling skill that agents load first. The order that follows is list_projects, then query_profile_types, then the profile tool. Polar Signals reports the model picks the right tool far more reliably with that sequence in context.

3. 3

   Expect the skill to cost context on every investigation

   The vendor puts it at about 2,850 tokens, down from roughly 4,700.

4. 4

   Prefer sandwich and focus paths over full dumps

   Both new tools exist specifically to keep whole flame graphs out of the context window.

5. 5

   Know the truncation

   Sandwich output is capped at the top 100 frames. The cut is breadth-first, which means the tree is trimmed by depth, so the deepest row shown is not necessarily the true entry point.

Paste this into your agent’s instructions

Profile output is truncated by design, and the truncation is what stops it flooding the context. An agent that forgets this reports the visible top of a tree as the root.

When you answer using Polar Signals MCP tools, state:
- The profile type, the label filter and the time range.
- Which tool produced the numbers: the table, the sandwich
  or the flame graph.
- That sandwich output is capped at the top 100 frames, so
  the deepest frame shown may not be the true entry point.
Never state that data does not exist. State the query above
and the window it covered.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Only on Polar Signals Cloud, and only when you cannot already see the answer. The UI is the direct route, and clicking a flame graph is faster than describing one when you know where to look.

Self-hosted Parca keeps its own query API and web UI. Neither speaks MCP natively. The community davi17g/parcaprof-mcp bridges the gRPC API, at 0 stars.

For the questions profiling cannot answer, Grafana, Prometheus and Honeycomb sit next to it: profiling says which function is expensive, and those say which request failed.

pprof remains the standard tool for a single captured profile. Continuous profiling exists because the interesting profile is usually the one nobody thought to capture.

With MCP, the model navigates the flame graph instead of receiving it, which is the difference between a bounded answer and a context window full of frames.

Use it on Polar Signals Cloud. On self-hosted Parca, use the UI or vet the community server first.

Parca MCP server setup

The server is hosted by Polar Signals, so nothing is installed. Its endpoint answered HTTP 401 to an unauthenticated call on 3 October 2026, which is a live gated endpoint behaving correctly.

One condition decides which route applies. The vendor connector belongs to Polar Signals Cloud, and the Parca project has not published one for self-hosted installs as of 3 October 2026.

Both configurations below come from Polar Signals' own announcement post. Since 12 February 2026 the endpoint also accepts OAuth, so clients with a browser flow can skip the token.

claude mcp add --transport http polarsignals \
  https://api.polarsignals.com/api/mcp/ \
  --header "Authorization: Bearer <your-token>"

{"mcpServers":{"polarsignals":{
  "url":"https://api.polarsignals.com/api/mcp/",
  "headers":{
    "Authorization":"Bearer <your-token>"}}}}

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

curl -s -o /dev/null -w "%{http_code}" \
  https://api.polarsignals.com/api/mcp/     # 401

Read-only access, permissions and security

• Profiles carry function and file names

  A flame graph is a map of your source tree, including internal package and symbol names, and the source tool goes further by annotating cost onto code.

• The account scopes it

  Access follows the Polar Signals Cloud project, so what an agent can profile is what the credentials can already open.

• All eight tools read

  Nothing in the documented set writes to a profile store or changes a project.

• Output is bounded by design

  The 100-frame sandwich cap and the flame graph focus_path limit how much reaches the model, which limits exposure as a side effect of limiting cost.

Troubleshooting

You cannot find an MCP server in the Parca repository

Expected. The Parca project has not published one as of 3 October 2026. The vendor server belongs to Polar Signals Cloud, at api.polarsignals.com/api/mcp.

A tool name from an older article does not exist

Six names changed on 30 June 2026. profile_types, labels, values, get_profile and query_source_report became query_profile_types, query_profile_labels, query_profile_label_values, query_profile_table and query_profile_source.

The flame graph tool returns a table

That was the old behavior, which Polar Signals described as the same naming sloppiness. query_profile_flamegraph returns the real hierarchy.

The deepest frame is not where the call started

Expected. The sandwich caller tree is truncated to the top 100 frames breadth-first, and the tool output says so: the true entry point may lie beyond what is shown.

A frame has no function name

Unsymbolized frames return a hex memory address instead, and focus_path accepts those addresses, so kernel and native stacks stay navigable.

Parca MCP server: Reference

What the vendor documents

Parca has 4,987 stars, is Apache-2.0 and was pushed on 29 September 2026. Public discussion of the connector itself is thin, so the vendor's own release writing is the useful record, which is unusually specific about cost and about its own mistakes.

“As you can see the tools were named quite randomly.”

“To reason about structure from a flat table, the model has to load the whole table into its context and reconstruct the relationships in its head. The interesting function is three rows from the top, but the model is carrying all several hundred rows to find it.”

“A skill is context too, of course, loaded on every investigation. [The rewrite is] about 40% smaller, from roughly 4,700 tokens down to 2,850.”

A GitHub repository search for parca and mcp on 3 October 2026 returns one Parca server, davi17g/parcaprof-mcp, at 0 stars. Neither the parca-dev nor the polarsignals organization holds an MCP repository. Tool names, the frame cap and the skill sizes are quoted from the Polar Signals release posts of 17 July 2025 and 30 June 2026.

Should you connect an agent to Parca?

Yes. Best for table first, then sandwich into the suspicious function, then source.

Strongest fit

Polar Signals Cloud users chasing a CPU or memory regression they cannot yet name.

Main advantage

The tools were designed against the context window, not adapted to it. Sandwich trees and focus paths keep whole flame graphs out of the answer.

Main weakness

It is not available to self-hosted Parca, which is the version most people run.

Operational risk

Truncated caller trees read as complete, so a reported entry point may simply be the deepest frame that fitted.

Frequently asked questions