Code Review Graph MCP

by tirth8205

An MCP server that benchmarks 38x-528x token reduction on code review. It shipped v2.3.5 three days ago. 17,500+ GitHub stars. Built on tree-sitter. Runs 100% locally. tirth8205/code-review-graph is the local-first code intelligence MCP server purpose-built for code review. MIT license. Python (91.2%). Latest release v2.3.5 published May 25, 2026 — three days before this listing. Tag-line in the repo: 'Real-time token savings, visible to humans.' What 30 Specialized Tools Get You Most MCP servers expose three or four tools. Code Review Graph exposes 30, each tuned for a specific review workflow. The headliners: Blast-radius analysis — given a proposed change, return every function, class, and file affected, ranked by coupling strength. Semantic search via vector embeddings — natural-language queries that traverse the graph instead of running grep across files. Impact detection — affected flows, test coverage gaps, downstream consumers of any changed symbol. Architecture analysis — auto-detected code communities, coupling patterns, dependency cycles, hotspot identification. Multi-repo support — register all your services and run cross-repo impact analysis on shared library changes. Real-time token usage display — the human reviewer sees exactly how much the AI is saving versus a naive full-repo prompt. The 30 tools are not bloat — they are the review-specific verbs the underlying agent needs to do PR review well. Blast-radius, coverage, coupling, dependency, hotspot, community — these are the questions a senior reviewer asks. The MCP exposes them as primitives. Why The 38x-528x Number Is Believable The token reduction range is wide on purpose. At the low end (38x), the review touches a substantial slice of the codebase and the agent still has to read meaningful context. At the high end (528x), the review is a localized change in a large monorepo and the agent gets to skip 99.8% of what it would otherwise read. The mechanism is simple. A naive prompt is 'here is the diff, here is the whole repo for context, what should I worry about.' That blows the token budget on every review. The MCP-driven prompt is 'here is the diff, the MCP says these 12 functions are affected and these 4 tests cover them, focus there.' The agent's context drops by orders of magnitude and the review quality usually improves because the agent is not skimming noise. The Money Math For A Team For a 10-developer team running roughly 50 PRs per week through Claude Code review, the naive token spend is in the $1,200-1,800/month range — full-repo context on every review compounds fast. A 38x reduction collapses that to roughly $30-50/month. A 528x reduction collapses it to under $5/month. Most real teams will land somewhere in the middle, but the cost-cut order of magnitude is consistent. The free MIT license means there is zero recurring infrastructure cost to capture those savings. The tree-sitter parser is open source. The graph index is local. The MCP tool layer is local. The model API spend you replace is real money that goes back into the budget. The Local-First Story For Regulated Codebases Tree-sitter is the right choice for this. It parses real code structure, not text, so the graph reflects actual call relationships and not regex approximations. And the entire pipeline — parser, indexer, MCP tools — runs on the developer's machine. No code or metadata is shipped to a vendor. For regulated codebases (defense, healthcare, finance, automotive) this is the line that decides whether AI code review is procurable. The local-first design is the audit story that closes the procurement loop. What To Do With It This Week If you run Claude Code, Cursor, or Codex on PR review in a codebase north of 100,000 lines, install it tonight. The setup is one command per client. If you are a tech lead approving the AI-coding-tools budget for a team of 10+, the 38x-528x reduction on review-specific spend is the line item that justifies the install in the next budget cycle. If you are on a regulated codebase that previously blocked AI tooling because of egress concerns, this is the week to revisit — tree-sitter, local, MIT, audit-ready. For broader context on the AI-coding-stack-getting-efficient pattern, see CodeGraph — the general-purpose code knowledge graph that pairs naturally with Code Review Graph as the broader-context layer — and Pi Coding Agent, the model-agnostic OSS CLI that gives you the harness to drive any of these tools. The pattern across the stack is the same — constrain what the AI is allowed to read, hand the human the steering wheel. Our breakdown this week of the AI-replacement narrative walks through why this is the right architecture.

FAQ