Govern the AI
your employees are building.
Your teams create skills, agents, and automations in Claude every day. You approve what becomes official. Everyone runs it, no new platform.
Understand knacks in 10 seconds. ↑ Send it.
Everyone's using AI. Nobody's in charge.
Skills, agents, methods: your employees are shaping how the company works in Claude, every day. None of it goes through anyone.
Anyone allowed to share can push their AI work to the whole organization. No review step exists.
The only alternative admins have is switching sharing off. So the methods that actually work stop spreading too.
Ten versions of the same method float around in docs and chats. Nobody knows which one the company stands behind.
And nobody can answer the simple question: what AI runs here, and who said yes to it?
One path: proposed, approved, official.
One person figures it out. Approve it. The whole team runs it in Claude. Nothing ships without the middle station, and only the chosen team receives it.
Skills and agents. Same path.
One registry for the AI your company runs on: who owns what, who approved what, which version is live.
Company skills
The methods your teams run in Claude: proposed by the person who proved one works, approved by the team lead, delivered to the right team, current by default.
What a company skill is →Agents
Agents act: they send, write, and change records. Before one acts for your company, it takes the same path: proposed by the person who built it, approved by the team lead, bound to the exact version and its scope of access.
How agent governance works →Same review, same registry, same rule: nothing ships unreviewed. Go deeper: Who approves an AI agent?
Three things you stop losing.
New joiners start from your best person's method on day one, not month three.
One approved way to do repeated work. When it improves, everyone improves.
People leave. Their approved methods stay: versioned, owned, running.
Run one of these in your first week.
Renewal call prep
Every CSM preps renewals the approved way: risk, usage, stakeholders, next steps.
Open the play → SalesEnterprise account research
Your best rep's research routine, run by every rep before a first call.
Open the play → SupportSupport escalation
Escalations that follow the same approved path, tone included.
Open the play → OperationsWeekly operating report
The report everyone actually reads, same structure every week.
Open the play →What exactly is a skill? The canonical definition.
The domain expert, not IT.
Sees what's approved, owned, and current.
Watches nobody: no chats, no screens, no run counts.
knacks for Heads of AI → CROApproves the best rep's method.
Every rep runs it in Claude, current by default.
knacks for CROs → COOControls the methods the operation runs on.
Reporting, triage, and analysis stop being tribal.
knacks for COOs → RevOpsShips only approved methods to the revenue team.
One version, one owner, one changelog.
knacks for RevOps →Built to pass your toughest reviewer.
Your skill library lives as plain files in a GitHub repository you own. We connect through a GitHub App scoped to that single repository, nothing else.
Ephemeral access tokens, revocable in one click. No long-lived credentials stored.
Employees don't need GitHub accounts. The app records approved changes on their behalf.
Bring your own Anthropic key. Generation runs on your key, under your data agreement.
No capture and no usage tracking, ever. Publishing is a deliberate act, and knacks never sees chats, screens, or how often anyone runs a skill.
Approvals bind to the exact version reviewed. If a skill changes afterward, the previous approval no longer applies. There is no bulk approve.
Exportable by design: your skills are markdown files in your own repository. They never leave your side.
Straight answers.
What exactly is a skill?+
A small, approved, versioned piece of know-how: the method in plain language, the files it needs, examples, tests, and a named owner, in the open format Claude reads natively. It doesn't describe the work; it does it.
Do you watch what employees do in AI tools?+
No. knacks has no capture and no usage tracking: no access to anyone's chats or screen, and no record of who runs what. Nothing enters the system unless someone chooses to publish it.
Is this another platform we have to roll out?+
No. Employees publish from Claude, team leads approve from Claude, teams run skills in Claude. There is no knacks app to live in and nothing for your people to learn. Admins touch a small setup area a few minutes a month.
Our teams use more than Claude. Why Claude-first?+
The chaos is multi-tool; the memory shouldn't be. We ship where skills run natively today: Claude, across web, desktop, and Claude Code. What you publish is open markdown you own: portable by design, locked to no one, including us.
Couldn't our engineers just set this up themselves?+
If your whole company is engineers, genuinely, try it. knacks exists for everyone else: a publish flow non-technical people actually use, approval that belongs to team leads, and one library that keeps everyone on the latest version.
Is this another chatbot?+
No. knacks doesn't add a chat interface. It is the publishing, approval, and distribution layer for the AI work your teams already do, delivered into the tools they already use.
Do we have to replace our existing tools?+
No. Skills run in Claude and live as files you own. Nothing proprietary holds your knowledge hostage.
Do you govern agents too?+
Yes. Agents take the same path as skills: proposed by the person who built one, approved by the team lead, bound to the exact version and its scope of access. Nothing ships unreviewed.
How fast can we start?+
One team, one repeated task, one approved skill, typically live in the first week. We're service-led at the start on purpose: you get live skills, not a strategy deck. See how we start.
You have already paid for this.
Your people found the good methods months ago. You just have nothing to show for it.
The brief your team reinvented this week? Someone in the building solved it in March.
The doc everyone still copies has been wrong since the pricing change. It keeps answering under your brand.
The rep who left last month took her method with her. Nothing official survived her notice period.
Unreviewed sharing is not a phase you grow out of. It is how wrong methods become habits.
Every week without an approval layer, you pay for the same discoveries again. Stop paying twice.
Stop letting your best AI work die in private chats.
Book a walkthrough. Bring one team and one repeated task, leave with your first skill: published, approved, and in use. Everything from Claude, nothing to install.