Flameox MCP
Local profiling MCP server: an agent runs profilers and benchmarks and gets bounded evidence about hotspots
- 240 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Capture runs your process and profilers on the machine where the agent runs
- Preserved data may contain paths, environment variables and program output
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/flameoxDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Capture runs your process and profilers on the machine where the agent runs
- Preserved data may contain paths, environment variables and program output
This is third-party code. Review the repository files before installing.
What it does
Flameox coordinates profilers, benchmark tools, trace processors and direct local targets. It turns an explicit artifact or command into bounded evidence: it records observed inputs, execution provenance, typed evidence, coverage and limitations. The agent picks hypotheses and experiments while Flameox invents nothing. The scope is application and native code, GPU kernels and inference stacks. No workspace or config file is needed. Analysis takes exact artifact paths, and capture takes argv and an absolute working directory. Flameox does not search parent directories, edit project files or run as a service. Capture takes argv, not a shell string. Results arrive in pages with a `next_page` link and a continuation reads artifacts already collected and does not rerun the target. Preserved data gets a durable `evidence_id` shown by `inspect_evidence`.
Who it is for. Performance engineers who hunt bottlenecks in Python, native code and GPU kernels together with an agent and want verifiable data.
Good fit when
- A benchmark slowed after an edit and the agent captures a profile before and after and compares them
- A GPU kernel is slow and you need hotspot data from a profiler rather than guesses
- Profiling results are preserved with an `evidence_id` so you can return to them a week later
- A huge trace is read in pages through `next_page` without rerunning a heavy program
Not a fit when
- You need long monitoring of a live system: Flameox captures one target on an explicit command
- You need a hosted service with dashboards: the server is local and runs no hosting
- One execution timing is enough: the language's stock profiler is simpler
Example request
Run the benchmark python benchmark.py under a profiler, find the three hottest functions and show the evidence with sourcesMCP
- Transport
- stdio
- Authentication
- not required
Limitations
The server has no workspace binding: the artifact path and an absolute capture directory are given on every call. Capture runs your command, so the target runs on the machine with the agent's rights. Work is tied to the live request: cancellation applies at once and no background jobs survive a restart. Preserving results is optional. The MCP registry version is 0.2.9 and the license is MIT. Setup is done with `npx flameox@latest setup` and updates with the `update` command, which keeps each client's settings.
How to disable. Remove the Flameox registration in the client (`claude mcp remove flameox`) or the config edit made by `setup`, and the package if you installed it with `pip`.
FAQ
How do I connect Flameox?
Run `npx flameox@latest setup`: it detects supported clients and asks which to configure. For automation name clients explicitly, for example `--client codex --yes`, and `--dry-run` shows the actions without writing.
How do I start the server by hand?
With `uvx flameox mcp serve` over stdio. `flameox mcp inspect` shows the tool catalog and `--tool TOOL_NAME` prints exact schemas.
Which tools cover the data lifecycle?
`prepare_providers`, `preserve_evidence`, `rescue_evidence`, `query_evidence` and `inspect_evidence`, plus `analyze_` and `capture_and_` pairs for each operation.
Does Flameox rerun my program on continuation?
No, a continuation reads already collected artifacts. The target runs once unless you ask for an experiment with several cases.
How do I update an installed Flameox?
With `npx flameox@latest update`: `update --check` only reports available updates and `--version 0.2.8` selects a version, including for rollback.
Discussion
No questions or reviews yet.
Related
MCP serversA local code graph with version history: an MCP server that tells the agent who calls a function and what an edit breaks
An MCP server that builds a code graph once and serves the answer to the agent instead of re-reading files
agentburn reads local logs of Claude Code, Codex and other agents and shows which windows and tasks consume tokens and money.
An MCP server and CLI for code research: search and understand your project and all of GitHub with fewer tokens
Okto Pulse tracks ideas, specs, tasks and bugs on a gated board that agents operate through 285 MCP tools
DevLens analyzes a repo once and serves agents ranked files, call flow and change impact over MCP, no file-by-file reading
README in short
The text describes Flameox as a coordinator of profilers and benchmarks that returns bounded evidence instead of conclusions. Then come a quick start through `setup`, direct CLI commands, starting the server, an MCP section with paged output and links to documents on interfaces, safety and data storage.