Open source · tools for coding agents
We build tools for coding agents, then check whether they actually help.
Everything here is open. The tools ship as real binaries and extensions; the evaluations come with the raw transcripts, so you can read the runs and disagree with us. Have a poke around.
Tools you run
softwaregrove live
Byte-precise, structural access to a codebase for a coding agent, instead of reading whole files. One binary that speaks a CLI and MCP, with grammars loaded at runtime.
Forge
Turns a repository into a structured engineering knowledge base for Claude Code, then drives the plan → implement → review loop from it.
Forge CLI
The same Forge pipeline as a standalone command line, ported onto the pi coding agent as a TypeScript extension.
SkillForge
Reusable skill packs you drop into Claude Code or another agent harness, so a capability travels between projects instead of getting rebuilt each time.
Evidence you can check
benchmarksis-grep-enough live
Does an agent exploring a big codebase need more than grep? A blind-judged race of text search against grove and a language server, run across ten real repositories.
tokbench
An independent look at token-saving middleware on a real agentic SDLC workload. Pre-registered protocol, with every transcript published so the numbers are yours to re-check.
The rest is on GitHub — pi extensions, the grove grammar registry, and the testbenches that feed the numbers above.