recx

res extra commercium: things outside commerce. A way for agents to find reusable, permissively licensed code from solo and small-team projects.

A RECX.yaml at the root of a repository tells other people's coding agents what the repository does, which parts of it they can take on their own, and which checks show those parts work. recx is that format, a validator for it, and a Claude Code plugin that keeps the file current.

Why

Coding agents have increased the amount of software that individuals and small teams produce. Much of it remains internal tooling, and the projects that are published are hard to find among the number of new repositories. Evaluating an unfamiliar open-source project also takes time, so writing a similar tool is often the quicker option, and comparable tools get written more than once.

recx moves that search and evaluation to the agent doing the work, in two parts. The first is the manifest: a RECX.yaml in each participating repository that describes the project's capabilities, its separable modules and their interfaces, and the commands that verify them, in a fixed machine-readable form. The second is search: a way for an agent, while planning new work, to find manifests with relevant modules, check their evidence, and reuse a module rather than write it again, recording where it came from in its own project's manifest.

Taking part is meant to cost a repository little. One prompt produces a complete, valid manifest, and the Claude Code plugin checks it whenever a session ends, so it stays valid as the code changes. No hosted service is involved. The fixed filename and the GitHub topic recx serve as the index, and a scheduled workflow in this repository collects the valid manifests it finds into a static file each day. In GitHub's web interface, manifests can also be found by searching code for path:/(^|\/)recx\.ya?ml$/i; GitHub's code search API does not accept that regular expression.

The manifest

recx: "0.1"
name: tinyfmt
summary: Formats byte counts as human-readable strings like 1.2 MB.
license: MIT
status: stable
capabilities:
  - format a byte count as a human-readable size string

Those six keys are the minimum. The optional parts carry most of the information: modules with interfaces and extraction notes, evidence anyone can rerun, and records of what was reused from where. spec/SPEC.md defines every field, puts a length limit on every string, and sets out how to read a stranger's manifest safely.

Only permissively licensed projects qualify (MIT, Apache-2.0, the BSDs and similar), so anything listed can be reused without asking.

Get started

• Claude Code: /plugin marketplace add jbmv/recx, then /plugin install recx@recx (or, on Claude Code 2.1.275 or later, /plugin install recx --marketplace jbmv/recx). Then ask Claude to add a RECX.yaml. A Stop hook keeps the manifest valid; set RECX_STOP_HOOK=warn or off to soften or disable it.
• Any other agent: paste the prompt in ADOPT.md.
• By hand: start from the example above, then run uvx --from recx-cli recx check.

Then add the GitHub topic recx to the repository, so the daily index finds it promptly.

The plugin's hook runs when a session or one of its subagents stops (Claude Code's Stop and SubagentStop events). It blocks on an invalid manifest, and it reminds, without blocking, when files under a module's path changed during the session while RECX.yaml did not. A commit can record that module entries were checked against its tree with a trailer such as Recx-Checked: lib, cli; the reminder then disregards that commit's changes to those modules and the changes of the session's commits it descends from. RECX_DRIFT_REMINDER=off turns the reminder off. The manifest skill describes both in full.

Searching

recx search "retry an http request with exponential backoff"
recx search parse rss feeds --language python --json
recx show OWNER/NAME

recx search downloads the published index (https://jbmv.github.io/recx/v1/index.json), caches it for six hours, and ranks entries by keyword overlap with the query, weighting capabilities above summaries and descriptions. It returns every entry that shares a term with the query, ordered by score, and leaves the judgement of relevance to the agent reading the results. --json gives the structured form a tool or agent should use, and the same result is available from Python as recx_cli.search.search(). recx show prints one indexed entry in full. Neither command needs a GitHub token or runs anything from a repository.

Every result is written by the repository's author and has been checked only against the schema. The output says so in a fixed notice, and the JSON lists the author-written paths in untrusted. The index never contains a manifest's setup or run commands. Before reusing anything, read the repository's licence file and its current RECX.yaml, and run its evidence only in a sandbox. docs/integrating.md explains how to call search from another tool, and docs/design/search.md describes the index, the crawler and the ranking in full.

What's here

Status

Experimental. The spec is a v0.1 draft and may change before 1.0; every manifest states the version it was written against. recx check enforces the whole spec except rule 4, that the license file matches license; the search crawler applies it as far as GitHub's licence detection allows. Search is in its first version: it ranks by keywords, has no synonym handling, and has no MCP server yet.

Kyosei

I built recx for Kyosei, a record-keeping tool for running projects with coding agents. Kyosei isn't public yet. It will generate its own RECX.yaml from its records. Nothing in the spec depends on Kyosei: it is one client of the spec, like any other tool.

License

MIT