code-review-graph vs codegraph-mcp
Local-first code intelligence graph for AI coding tools: Tree-sitter AST graph + blast-radius analysis served over MCP, so agents read ~82x fewer tokens per review question. — versus — No-train, on-prem code knowledge graph served to AI agents over MCP — symbols, call edges, cross-language links and blast-radius queries, with a hash-chained audit log of every read.
Both pre-index a semantic code graph and serve it to agents over MCP; code-review-graph adds blast-radius/risk scoring and a CI review action, codegraph-mcp stays a leaner symbol/caller index.
| code-review-graph | codegraph-mcp | |
|---|---|---|
| Stars | 21k | 5 |
| Forks | 2.1k | 0 |
| Language | Python | Python |
| License | MIT | NOASSERTION |
| Last activity | yesterday | 15 days ago |
| Topics | coding | coding, rag |
| Curated connections | 4 | 6 |
code-review-graph — the curator's take
The rare benchmark-honest repo: it tells you the 528x number is the best case and its recall metric is circular. Install once, it configures 14 platforms (Codex, Claude Code, Cursor...). Reach for it on monorepos where review context is the token sink; skip on small repos where a grep costs less than the graph's own metadata.
codegraph-mcp — the curator's take
Niche but real: if agents must understand code that cannot leave your building AND compliance asks 'what exactly did the agent read', the tamper-evident audit chain is the only game in town; cross-language call edges (TS fetch → Go/Python handler) catch what one-file context misses. NOT for most teams yet: 5 stars, single vendor, and — deal-breaker until fixed — no clear open-source license (NOASSERTION on GitHub). Treat it as an evaluation candidate, not a dependency.