Multi-Agent Code Review MCP Server

An MCP (Model Context Protocol) server that runs a multi-agent code review pipeline built with LangGraph. Point any MCP host (Claude Desktop, Cursor, Claude Code) at it, hand it a git diff, and a graph of specialized LLM agents reviews the change for security, quality, and logic issues, then a supervisor agent aggregates their findings into a single pass/fail verdict.

This is not an API wrapper. It is a stateful multi-agent graph behind a production-style guardrail stack.

Why this exists

Most "AI code review" demos are a single prompt that says "review this code." Real review is adversarial and multi-perspective: a security reviewer, a quality reviewer, and a correctness reviewer each look for different things, and someone senior reconciles them. This project models exactly that as a LangGraph state graph, and exposes it over MCP so it plugs into the tools developers already use.

Architecture

          validate + rate-limit + audit        (deterministic guardrails)
                        │
                        ▼
                    prepare                      (redact secrets before any LLM call)
                        │
        ┌───────────────┼───────────────┐
        ▼               ▼               ▼
   security         quality          logic       (3 agents, run concurrently)
        └───────────────┼───────────────┘
                        ▼
                   supervisor                     (aggregate, gate on severity)
                        │
                        ▼
                  pass / fail verdict

• LangGraph manages shared state and the fan-out / fan-in of the three reviewer agents.
• Each agent returns structured output (a Pydantic AgentReport), so malformed model output is rejected at the boundary, never propagated.
• The supervisor merges agent findings with a deterministic secret scan and applies a configurable severity gate.

Industry-level features (not a student toy)

• Input validation — size limits, encoding checks, file-count guards.
• Deterministic secret pre-scan — regex catch for AWS keys, private keys, bearer/Slack tokens, API-key assignments. Runs before any LLM call.
• Secret redaction — matched secrets are redacted before the diff is ever sent to an external model.
• Rate limiting — in-memory token bucket per configured window.
• Audit logging — every call logged with a SHA-256 of the input (never the raw source), so the audit trail is not a secondary leak.
• Config-driven severity gate — a review fails only if a finding meets the configured threshold; fully env-overridable via pydantic-settings.
• Graceful degradation — if one agent's model output is unparseable, that agent is skipped, not the whole review.
• Tested — the deterministic core (security + supervisor logic) has a full pytest suite that runs with no API key.

Tech stack

Python · MCP (FastMCP) · LangGraph · LangChain · Pydantic · pytest

Setup

pip install -r requirements.txt
export CODEREVIEW_LLM_PROVIDER=anthropic
export ANTHROPIC_API_KEY=...        # or OPENAI_API_KEY with LLM_PROVIDER=openai

Run the tests (no API key needed)

pytest

Run the server

python -m src.server

Register with Claude Code / Cursor

claude mcp add code-review -- python -m src.server

Then in your MCP host, call the review_diff tool with a unified git diff.

Configuration

All settings are environment variables prefixed CODEREVIEW_:

Variable	Default	Meaning
CODEREVIEW_LLM_PROVIDER	anthropic	anthropic or openai
CODEREVIEW_FAIL_ON_SEVERITY	high	Gate: fail at/above this level
CODEREVIEW_MAX_DIFF_BYTES	100000	Reject larger diffs
CODEREVIEW_RATE_LIMIT_PER_MINUTE	20	Max reviews/minute

Note on MCP versioning

requirements.txt pins mcp>=1.28,<2 on purpose: MCP SDK v1.x is the stable production line, and v2 is a breaking rework. Pinning below v2 is the recommended practice for anything depending on mcp.