Skillpipe
Git-native CLI for syncing AI agent skills across environments.
Define your agent skills once in a GitHub repository, install them into Claude Code with a single command, and let agents propose improvements through Pull Requests.
The skills drift problem
AI agents need one source of truth for instructions across machines, projects and teammates. Without it, every laptop ends up with a slightly different copy.
- 01Machines diverge — every developer ends up with a different version of the same skill.
- 02Improvements don't propagate — a fix on one laptop never reaches the rest of the team.
- 03No audit trail — instructions executed by inference deserve the same review as code.
What Skillpipe gives you
Git-backed source of truth
Skills live in a GitHub repository your team already knows how to review and version.
One command to install or update
skillpipe install and skillpipe update keep every machine in sync with the repo.
Paranoid validation
Schema checks plus secret scanning for OpenAI, Anthropic, GitHub, AWS, Slack and Google before anything touches disk.
Pull-request only contributions
skillpipe propose opens a PR. Main is never written to directly — every change is reviewable.
Snapshots, not symlinks
Installed skills are copies pinned to a commit. Upstream changes never silently propagate.
Multi-target adapters
Claude Code today. Cursor, Codex and custom paths via a small adapter API.
Four commands. That's the whole flow.
Initialize
Scaffolds ~/.skillpipe/ and bootstraps the operator-manual skill so agents know how to use the tool.
skillpipe initConnect a repo
Point Skillpipe at the repository where your team's skills live.
skillpipe repo connect <url>Install a skill
Validates and writes the skill into Claude Code, or whichever target you've configured.
skillpipe install <name>Stay in sync
Pulls the latest versions, lockfile-tracked. Pair with --dry-run to preview every change first.
skillpipe update
Security model
Skills are instructions executed by inference. Skillpipe treats them like code that runs.
Secrets caught before install
Provider keys for OpenAI, Anthropic, GitHub, AWS, Slack and Google are flagged before they reach the lockfile.
Dangerous patterns flagged
rm -rf, curl-pipe-sh and other risky shell idioms are surfaced during validation, not after they've run.
Main is PR-only
Every change to a skill is a Pull Request. Reviewable, revertable, traceable.
Built for AI agents
Skillpipe ships with a bundled skillpipe-cli skill — the operator's manual the agent reads to learn the tool: mental model, prerequisites, error codes, hard rules and anti-patterns. The skillpipe propose flow then lets agents improve their own instructions by opening a PR, never by pushing to main.
Open source. Built by Levante.
Skillpipe is MIT-licensed. We build it in the open with the same community that uses Levante.
Made by Levante