ripwire

Coding agents burn an enormous share of their context window just orienting themselves in a codebase they have never seen before: reading whole files end to end, grepping the same terms repeatedly, and re-deriving call relationships that an experienced engineer already carries in their head. ripwire, built by Red Hat's Emerging Technologies group, treats that orientation problem the way ripgrep treats text search: something a fast, dependency-free binary should compute once and hand back as structured data, rather than something an agent should rediscover the slow way through trial, error and repeated file reads that quietly eat the token budget.

Point ripwire at a repository and it builds ranked call graphs annotated with complexity and churn, an impact analysis showing the blast radius of a proposed change, and a quality panel scored from six independent evidence families (structural, lexical, confusion, historical, colocation and state). It covers 24+ languages through vendored tree-sitter grammars, needs no network access and no external services, and outputs XML by default alongside JSON, HTML graphs or an MCP interface. Ready-made skills wire it directly into Claude Code, Codex, Cursor, Windsurf and Gemini, so an agent can call it as a tool rather than a human running it manually first.

It fits teams running coding agents against real, moderately large codebases, where the token cost of an agent re-reading files on every task is a measurable expense, and where a structured map of what calls what would replace several rounds of grep-and-read entirely. Because it ships as a single offline binary with prebuilt macOS and Linux releases, it also fits security-conscious or air-gapped setups that cannot send source code to a hosted context service. The official Red Hat backing (178 repos under the redhat-et org, ripwire as one of its pinned projects) gives it more institutional accountability than a typical solo-maintainer tool in this space.

The headline efficiency numbers - indexing this very repository in 0.25 seconds and 6.6 MB versus a named competitor's 46.8 seconds and 391 MB, retrieving only 5.0% of the tokens a grep-and-read approach spends, hitting 58.3% strict file@10 on retrieval - are all the project's own benchmark figures rather than independently reproduced results, and should be read as a claim, not a fact. The tool is pre-1.0 (v0.6.2) with weekly point releases, so command flags and output formats are still moving, and building from source requires a C++23 toolchain and CMake 3.24+ even though prebuilt binaries cover the common platforms.

The verdict: ripwire is a strong pick if your coding agents are spending real token budget rediscovering a codebase's structure on every task, and you want that fixed with a fast, offline, zero-dependency tool from an accountable source rather than a cloud service. The official Red Hat org and active weekly release cadence are genuine trust signals. Treat the specific performance numbers as project-reported rather than verified, and expect the CLI surface to keep shifting until it reaches 1.0, so pin a version if you wire it into an agent pipeline today.