Repo context at 6.6 MB: the anti-graph database

What it does ripwire is a zero-dependency C++23 binary that ingests a repository and emits a ranked, deterministic call graph. It tells coding agents what symbols to touch, what might break, and which tests matter—replacing the usual pattern of grepping for names and reading whole files. The README advertises signatures at 80% fewer bytes than bodies, and a same-answer methodology backs token savings that range from roughly 3× to over 300× against naive passes.

The interesting bit The project wears its homework openly: it catalogs 42 repositories and 67 papers it learned from, plus 237 tools it surveyed and dismissed, all in a lineage file that is automatically checked against the README on every run. It also publishes honest counterexamples where its own ranked bundles would cost more than reading the raw file, switching to whole-file mode instead. That transparency is unusual for a tool selling speed.

Key highlights

• Zero runtime dependencies; one binary that runs fully offline.
• Indexes its own repository in 0.25 s and 6.6 MB, versus 46.8 s and 391 MB for an unnamed graph-database MCP server it was benchmarked against.
• Supports a wide language list including Rust, C++, Python, Go, Swift, TypeScript, Java, and others.
• Optional MCP server interface for agents, though the CLI is pitched as the cheaper path because MCP verb schemas sit in context permanently.
• Same-answer verification: every token-saving row in the evaluation table is checked for correctness, not just byte count.

Caveats

• The compounding savings claim for multi-agent orchestrators is explicitly pre-registered but unrun; the README admits this has not been measured.
• The benchmark compares against a single unnamed “leading graph-database code-context MCP server”; no broader suite of competitors is shown.
• Warm query latency and index size figures come from a narrow set of repos (django, webpack, and itself).

Verdict Worth a look if you are feeding large or unfamiliar repositories to LLM agents and paying per token. Skip it if you already have a fast, cached indexing pipeline you trust, or if your agent workflow rarely starts cold on new code.