TokenSavevscode-review-graph

At a glance

Where code-review-graph is the better tool

• Multi-repository work

  A registry of repos and a cross-repo search tool. TokenSave operates on one project at a time — you can point it at another graph root, but there is no single query spanning all of them. If your work is genuinely multi-repo, this is decisive.

• Execution flows

  Call chains from entry points, assembled into named flows sorted by weighted criticality. TokenSave has the raw traversal — callers, callees, impact, call_chain — but never assembles them into ranked flows. Their own numbers put flow detection around 33% recall, so treat it as a lead generator rather than ground truth.

• Leiden community detection

  Proper clustering of tightly-coupled code, with recursive splitting for large graphs, plus wiki pages generated from the community structure. TokenSave's coupling and DSM analysis gets near this but never names the clusters.

• Framework-aware extraction

  Laravel/Composer PSR-4 resolution for PHP and Spring dependency-injection wiring for Java. TokenSave's extractors are generic per-grammar and will miss framework-mediated edges that no AST node spells out.

• Published, adversarial benchmarks

  ~63x median per-question token reduction across six repos and 0.69 average F1 for impact analysis against graph-derived ground truth. Publishing an F1 below 1.0 for your own core feature is the mark of an honest benchmark, and TokenSave publishes no comparable accuracy figure.

• A community an order of magnitude larger

  Roughly 31,500 stars against roughly 640. More contributors, more issues found, more likely to outlive its author's interest.

Where TokenSave is the better tool

• It can sit in the critical path

  A complete TokenSave invocation is about a quarter of a second with no interpreter to start. That is what lets it install a PreToolUse hook that fires on every Agent, Grep and Bash call and redirect symbol-shaped searches into the graph. Pay a Python cold start on every tool call and that design stops being viable — which is why code-review-graph is invoked deliberately rather than reflexively.

• Depth of analysis per symbol

  Cyclomatic and cognitive complexity, Halstead metrics, maintainability index, CRAP score, type hierarchies, inheritance depth, recursion detection, dead code, unused imports. code-review-graph's quality surface is essentially find_large_functions.

• Multi-branch and worktrees

  Per-branch databases, near-instant to create, with cross-branch search and diff. This is the thing that bites hardest when several agents work in parallel worktrees, and code-review-graph has no answer to it.

• Broader language coverage with deep extractors

  60 languages including Nix derivation fields, Protobuf messages and services, COBOL, Fortran, shader languages and legacy BASIC dialects — plus annotation extraction across 13 languages.

• Agent breadth

  12+ agents wired with native config formats, hooks, steering files and permissions.

Which should you use

Method

• 87 tools — asserted by test_tool_definitions_complete in src/mcp/tools/handlers/mod.rs; 86 when the optional ast-grep binary is absent from PATH.
• 60 languages — counted from the [features] tiers in Cargo.toml: 11 lite, 9 medium, 40 full. The README's "50+" is conservative.
• 0.24s — wall-clock real time for a complete tokensave status invocation, best of repeated runs on an Apple Silicon Mac. Your hardware and repo size will differ.
• Zero runtime dependencies — otool -L on the released binary; system frameworks only.
• code-review-graph figures — PyPI v2.3.8, the GitHub repository metadata and the project README, September 2026. Its benchmark numbers are its own and were not independently reproduced here.
• Where this page is wrong or has gone stale, open an issue and it will be corrected.