agent skill library · a protocol, not a platform
Same rules.
Every agent.
Sieve is a CLI that installs one skills-and-rules file every coding agent reads the same way — Claude Code, Codex, Cursor, and 20 more agents besides.
Two short questions. Then it pulls only the skills that fit from the registry, not the whole catalog. No registry? It falls back to the full bundled set on its own.
Stays honest over time
Catalog health, activity, skill usage — queryable via a self-hosted API, no account, no telemetry. On the command line, the same report:
A report, not a mutation. Run it on a cadence. Triage what it flags yourself.
Why not just paste rules into every tool?
The alternative is doing this by hand, per agent, forever. Or trusting each vendor’s own memory feature, and relearning it every time you switch tools.
| Sieve | Hand-copied rules | Vendor-native memory | |
|---|---|---|---|
| Works the same across agents | one AGENTS.md | copy/paste per tool | locked to one vendor |
| Plain files you own | yes | yes | vendor-held state |
| Skills fit the project, not everything | registry-recommended shortlist | all or nothing, by hand | no scoped-skill concept |
| Catalog health you can check | sieve check | nobody's tracking it | opaque |
| Portable if you switch agents | yes, by design | start over | start over |
Reads natively, or bridges in one line
AGENTS.md is read natively by most of these. For the rest, sieve bridge writes a thin pointer file back to it. Never a second copy of the rules.
How it stays honest
One command, tailored skills
Two short questions: new idea or existing project, any focus areas. Only the skills that fit get installed, plus the always-on guardrails. Not the whole catalog.
One protocol, every agent
AGENTS.md is the single source of truth. sieve bridge writes thin pointer files for the rest. You never maintain the same rules twice.
Six-pillar catalog
Planning, testing, review, debugging, verification, maintenance. Every skill has a clear job. The catalog stays legible as it grows.
Grows by use, not by design
Skills are proposed, validated, and promoted as real work surfaces them. sieve check keeps the catalog honest over time.
A starting point, not a ceiling
Add or drop skills anytime with sieve add and sieve remove, or see the current set with sieve list. The agent can also propose going beyond it mid-task. It flags that and asks first.
Portable, no lock-in
Every rule lives in a plain file you commit to the repo. Switch coding agents anytime — nothing to export, no vendor memory to rebuild.
Start in one command
No account, no server round-trip required, no lock-in.