Ship curated skills
to your whole team.
Authored by your engineers. Used by everyone. Never polluted.
The vision: a few engineers build the skills; a lot of people use them. ax will sync your blessed skills down to every laptop, read-only — and when a consumer’s agent hits an edge case, send a fix back upstream for you to review. We’re building this with a handful of design partners. Skills install, the local improve loop, and profile publishing all work today.
When your experts aren’t the ones who write the skills.
Lawyers, analysts, consultants — expensive professionals running agents your team built. git is the wrong altitude for them, and right now you’re hand-rolling repo locks to keep them from breaking it.
Curated, always current
Consumers receive the blessed set, synced read-only. No stale copies, no “which version are you on?” — and nothing they can accidentally clobber.
Can’t be polluted
No more locking the repo by hand. Consumers don’t push — their agents propose, your engineers approve. The gate does the governing.
Fixes flow back
Every real-world failure becomes a ranked suggestion — so the skill your whole team relies on gets better from the way they actually use it.
What works today vs what we’re building with partners.
Honest about the line: the local half of this loop is shipped. The sync fabric that turns it into a team registry is the design-partner work.
works today
- Skills install (npx skills add Necmttn/ax)
- Local improve loop (mine mistakes → reviewed fix)
- Profile publishing (ax profile publish → gist, consent-gated)
- Hooks SDK (typed Effect hooks, Claude Code + Codex)
building with design partners
- Sync-down (blessed skills read-only on every laptop)
- Suggest-up (agents propose fixes upstream)
- Review queue (ranked by real usage)
- Regulated mode (consent gate, provenance, fail-closed)
Fix flowed back: clause-numbering edge case
- seats hit
- 9
- failures, 14d
- 23
- repro
- synthetic, PII-free
- matter
- never sent
Mock of the review queue we’re building. Numbers illustrative.
One clean repo, made smart by how it’s used.
git versions the bytes. ax knows why each change happened and whether it worked — and ranks the noise so your authors only see what matters.
Author
Engineers write & bless skills in the repo you already use.
Sync down
ax will land blessed skills read-only on every consumer’s laptop, current.
Use
A consumer’s agent runs the skill and hits a real edge case.
Suggest up
The agent will send back a PII-redacted repro test — the failure, not the matter.
Review
ax will rank suggestions by real usage; your engineer accepts, it re-syncs.
The fix travels. The matter never does.
The thing that makes a suggestion useful is the sensitive context — privileged, for a law firm. So the local agent becomes the proxy: it ships a synthetic, runnable repro test, never the real input.
light mode · default
- Best-effort redaction, auto-share
- Failure-shape + synthetic repro test
- Frictionless for non-regulated teams
- A regression suite, built from real use
regulated mode
- Adversarial recover-pass — fail closed
- Local consent gate before anything leaves
- Provenance stamp — auditable, revocable
- Built for privilege, PII and compliance
Arm your whole team. Keep the repo clean.
We’re onboarding a handful of applied-AI teams as design partners. If you ship skills to people who aren’t engineers, let’s talk.