Before your AI builds it, check what already exists.

Every feature your agent writes from scratch costs tokens, time and new bugs. RepoHunter finds the maintained open-source version first and flags what to check before you install it.

claude mcp add repohunter -- uvx --from git+https://github.com/meetziggy/[email protected] repohunter-mcp
Your agent can now check what exists before it builds or installs.Needs uv · all install options
my-app · agent session
You

Write a hook that trims long command output. My Claude bill keeps climbing.

Agent

Checking what exists before writing anything.

repohunter.evaluate_repo("rtk-ai/rtk")
verdict
GO 89/100
status
maintained · last push 2026-10-03
license
Apache-2.0
fit
runs easily · compiled, small footprint
Agent

rtk already does this. Installing it with brew install rtk instead of writing one.

Examples:

Drop a GitHub link from any tab here, or paste one anywhere on the page. A quick read from public GitHub data; it doesn't know your project.

The rtk output above is real, captured 2026-10-04. The conversation is illustrative.

What RepoHunter does

It sits between your agent's plan and the install. One question, asked every time: does a good version of this already exist, and is it safe to bring in?

Where RepoHunter sits in an AI coding stack. You ask a coding agent such as Claude Code, Cursor or Codex for something. Before it builds or installs, it makes one call to RepoHunter, which runs locally: find_repos checks whether it already exists, evaluate_repo checks maintenance, license and fit, and the CLI scan looks for injection, hidden text, piped installs and leaked secrets. The verdict is GO (reuse it, pinned), MAYBE (check first) or SKIP (build or pick another). Guardrails: read-only, runs locally, token stays on your machine, pinned installs, about 1k tokens of context, untrusted text never fed to your agent. Measured: 56.0k tokens to build and fix versus 28.6k to reuse.

Before your agent writes a component

Reuse beats regenerating once the job is real.

A from-scratch Markdown renderer passed 5–6 of 16 everyday cases; the reused library passed 16. Field Notes 01

Before you install agent tooling

Plugins, skills and MCP servers come from strangers.

rtk: GO, but its README installs by piping a script into your shell; use Homebrew instead.

Before you ship something you sell

Some popular repos aren't licensed for that.

context-mode: GitHub can't identify its license; the file is Elastic-2.0, source-available.

Researching a topic 0.3.0

Papers and free books, as a cited reading list.

arXiv and OpenAlex papers, plus public-domain books from Open Library and Project Gutenberg. Ships in the next release.

I tried to prove it saves tokens.

On small jobs, it didn't. On a real component, building it yourself cost about twice the tokens, and still did less. 23 agent runs, same model and tests, raw data included.

Read Field Notes 01
56.0k → 28.6k
tokens: built then fixed, vs reused a library (Markdown renderer, median)
5–6 → 16
of 16 unseen everyday cases passed

Running agents on a team? I'll review what they depend on and load. See the audit or talk to me.

Why use RepoHunter instead of ChatGPT or Claude Code?

Use your assistant to reason and build; use RepoHunter for an adoption decision someone else can review. The full answer, including what isn't proven yet:

Use a general AI assistant to reason, explain, and help you build. Use RepoHunter when you need an open-source adoption decision tied to inspectable repository evidence, a consistent method, and a result someone else can review. They work well together: RepoHunter gathers and structures the evidence; you and your assistant decide what that evidence means for your project.

RepoHunter is read-only, shows the signals behind its guidance, and runs locally so you stay in control of your workflow. Missing or stale evidence should remain visible rather than being filled in with a confident guess. Its GO / MAYBE / SKIP result is a heuristic for a use case—not a security audit, a guarantee, or a verdict on a maintainer.

We do not yet claim RepoHunter is cheaper or faster. There is no controlled RepoHunter-versus-ChatGPT-or-Claude-Code benchmark yet. A defensible comparison must publish dated runs with wall-clock time, provider-reported tokens, tool calls, current prices, evidence completeness, and repeated-run consistency. Until then, the measured benefit is still an open question; the immediate value is a more reproducible, reviewable decision process.