About Fastmcp Pr Review
Fastmcp Pr Review is an MCP server published by strawgate in the Search category: fastMCP PR review server — three examples of increasing complexity (structured output → tool calling → production pipeline).
Six months or more without a commit doesn't mean the server is broken, but check the open issues before depending on it in production.
Install
git clone https://github.com/strawgate/fastmcp-pr-reviewThis server has no ConduID identity, so agent calls to it are not receipted. Pin the version you install and review the source before granting it credentials.
Tools (2)
review_pr_fast`fast.py`
Parameters not published.
review_pr_thorough`thorough.py`
Parameters not published.
Ask AI
Ask AI about Fastmcp Pr Review
Powered by Claude · Grounded in docs
Security checks
- ✓README presentThe repository ships documentation.
5 checks have not been run yet.
README
FastMCP PR Review
A GitHub PR review server built with FastMCP, with two review modes: fast and thorough.
Each mode is self-contained — read either file top-to-bottom to understand the complete pattern.
Review Modes
| Tool | File | What it does |
|---|---|---|
review_pr_fast |
fast.py |
One LLM call, structured output, no tools |
review_pr_thorough |
thorough.py |
Multi-pass pipeline: filter → review → verify |
Fast
One ctx.sample() call with result_type=PRReviewResult. Sends the full diff, gets back a structured review. Best for small PRs or quick checks.
Thorough
Multi-pass pipeline with batched file filtering, per-file review with a verification protocol, and agentic verification where the LLM explores the repo to confirm or disprove findings. Configurable intensity (conservative/balanced/aggressive). Only confirmed findings survive.
Quick Start
uv sync
cp .env.example .env
# edit .env with your GitHub and Gemini credentials
uv run fastmcp-pr-review
# or override runtime settings explicitly
uv run fastmcp-pr-review --gemini-model gemini-2.5-pro
make server-stdio
make server-http HOST=0.0.0.0 PORT=8000
The server auto-loads .env when started from the repo directory, and the Click launcher also reads GITHUB_TOKEN and GEMINI_API_KEY directly from the process environment. That means hosted environments like Horizon can inject those vars without extra CLI flags.
The server also exposes data tools: get_pr_info, get_pr_diff, get_pr_files.
Deploy to Horizon
For Horizon-style deployments that import a module and look for an mcp object, point it at the repo-root run_horizon_server.py. That file loads env vars and exposes:
mcp = create_server()
Set GITHUB_TOKEN and GEMINI_API_KEY in Horizon, and it can import run_horizon_server.py directly.
Add to Claude
Add the server to your Claude MCP config with a stdio entry that runs this repo through uv:
{
"mcpServers": {
"fastmcp-pr-review": {
"type": "stdio",
"command": "uv",
"args": [
"--directory",
"/absolute/path/to/fastmcp-pr-review",
"run",
"fastmcp-pr-review",
"--gemini-model",
"gemini-2.5-pro"
],
"env": {}
}
}
}
Keep GITHUB_TOKEN and GEMINI_API_KEY in /absolute/path/to/fastmcp-pr-review/.env or inject them via the host environment so Claude does not need inline secrets in its MCP config.
Use CLI flags for non-secret runtime settings like the fallback Gemini model.
Makefile
The repo includes a Makefile for the common local workflows:
make server-stdio
make server-http HOST=0.0.0.0 PORT=8000
make format
make lint
make lint-fix
make typecheck
make test
make build
Development
See DEVELOPING.md for testing and sampling patterns, and docs/ARCHITECTURE.md for architecture and design decisions.
License
MIT
README copy captured when this server was indexed; it may lag the repository. The original is authoritative.