Built a production code review agent in 2 weeks using MCP + ReAct.

The setup

A small dev-tools team wanted to add AI code review to their GitHub integration. They had:

• One full-time engineer (part-time on this).
• Three weeks before a planned launch.
• Strict latency target: under 15 seconds per PR for the median repo.
• A hard “no platform dependency” rule from leadership.

What they built

A ReAct agent against the MCP filesystem + git servers, with a custom post_review_comment tool that pushed structured comments back to GitHub.

from agentmatic.tools.mcp import MCPClient
from agentmatic.prebuilt import create_react_agent

fs = MCPClient.stdio(["npx", "-y", "@modelcontextprotocol/server-filesystem", repo_path])
git = MCPClient.stdio(["uvx", "mcp-server-git", "--repository", repo_path])

tools = (
    await fs.list_tools_as_agentmatic()
    + await git.list_tools_as_agentmatic()
    + [post_review_comment]
)

reviewer = create_react_agent(
    llm=Anthropic("claude-3-5-sonnet"),
    tools=tools,
    max_iterations=18,
)

Why this saved them weeks

They didn’t write filesystem or git tools. The MCP servers from the official @modelcontextprotocol org are already battle-tested by Claude Desktop and Cursor — same tools, same surface, no implementation cost.

Day-one functionality came from:

Component	Source	Lines of code they wrote
Filesystem tools	MCP server	0
Git tools	MCP server	0
ReAct loop	create_react_agent	0
GitHub comment tool	their code	~120
Eval harness	their code	~200
API + webhook plumbing	their code	~400

Numbers from the first month

• 8,400 PRs reviewed.
• 95% under 15 seconds; p99 under 32 seconds.
• 78% of comments accepted by human reviewers (vs 60% target).
• $0.04 average LLM cost per review.

Their take

“MCP solved the hard part of agent tooling for us. We didn’t write a single line of file or git access code. We focused on the comment quality, which was the only thing that mattered for shipping.”

— Founding engineer