The control plane for your organisation’s agent skills.
Publish the skills your org approves, install them into 17 AI agents, and block banned ones at run time in Claude Code — with a complete, immutable audit trail.
See every skill your agents already have and which ones are risky. Runs locally — no account, nothing sent.
Not ready to sign up? Check any public skill for free — see what it runs before your agents do.
Approved skills, installed into 17 agents.
A deterministic gate. No model on the path.
The gate evaluates a skill against the signed policy bundle on the machine — allow, warn, or block — with no network call and no model in the decision. In Claude Code, hooks ask it before every skill and subagent runs, so a banned one is refused. In the other agents, banned skills are removed on the next pull.
Policies are code, not prompts. Same skill, same bundle, same answer, every run.
From a script or CI. In Claude Code the hook answers with a denied tool call instead, with your policy's reason shown to the agent.
skillsvault pull --all installs approved skills into every agent it finds and removes banned ones.
Everything your skills need. Nothing they don't.
A registry, a gate, an audit log and a review step — designed to be boring in all the right ways.
Every published version is immutable and content-hashed. The store of record for your org.
Every policy bundle is Ed25519-signed. Each machine pins the signing key when it connects and rejects a bundle that doesn't verify.
Decisions are made on the machine. We receive the verdict for your audit log and the list of installed skills — never prompts, and skill files only when you publish or import them.
Which skill was checked, when, on which machine, and against which policy. Hash-chained, append-only, exportable.
| time | machine | skill | verdict |
|---|---|---|---|
| 10:32:04 | ana-mbp | acme/pdf-export | allow |
| 10:32:11 | ci-runner | acme/sql-runner | allow |
| 10:32:18 | ana-mbp | acme/teamleader-crm | block |
| 10:32:24 | kai-linux | acme/email-send | warn |
Any developer can propose a skill or a change. Governance sees the diff, what it does, and what the gate would decide, then approves or sends it back.
- acme/invoice-triagekai · new skillgate: allowedAwaiting review
- acme/deployana · change to v1.4.0gate: blocked by no-shellChanges requested
Blocks, warnings and new submissions in Slack, Teams or Discord, or as signed JSON for your SIEM.
A block links to a request. A reviewer exempts that one skill from that one policy for 7, 30 or 90 days, and every machine honours it until it lapses.
Claude Code and Gemini CLI install it as a plugin. MCP clients that support sign-in connect with just the server URL: people sign in, pick their organisation, and every skill they load is audited under their name.
Mapped against the OWASP Agentic Skills Top 10 — what skillsvault covers, and what it leaves to your platform.
The compliance system of record — owned by you.
The gate decides locally, from a signed bundle. Each decision is then recorded in your org's audit log: which skill, when, on which machine, against which policy. Rows are append-only and SHA-256 hash-chained, so a gap or an edit shows. Prompts are never sent, and skill files only when you publish or import them.
{
"occurred_at": "2026-06-25T10:32:18Z",
"machine_name": "ana-mbp",
"harness": "claude_code",
"skill_name": "acme/teamleader-crm",
"skill_version": "1.0.3",
"decision": "blocked",
"policy_name": "retire-teamleader",
"seq": 4182,
"prev_hash": "9c1e…a07b",
"entry_hash": "4f3d…e912"
}Start free. Scale to your org.
Govern your own agents for free. Pay per seat when your team shares one set of approved skills.
Open
For one developer governing their own agents.
- Versioned skill registry
- Install into 17 AI agents
- Runtime gate + signed policy bundles
- 1 member, 1 managed namespace
- 7-day audit history
- Alerts to Slack, Teams, Discord or a webhook
- AI authoring & review: 50 assists, 20 reviews a day
Team
For a team that shares one set of approved skills.
- Everything in Open
- Unlimited members, billed per seat
- Roles: owner, admin, governance officer, member
- Members propose skills, governance approves
- 3 managed namespaces
- 30-day audit history
- AI authoring & review: 200 assists, 200 reviews a day
Business
For the org that has to prove what its agents ran.
- Everything in Team
- Unlimited managed namespaces
- Unlimited audit retention
- Tamper-evident audit export (CSV / JSON) and a printable evidence report
- AI authoring & review: 1,000 assists, 1,000 reviews a day
- Priority email support
Need SSO / SAML, SCIM or a custom contract? Talk to us about Enterprise.
Audit history beyond your plan's window is kept and becomes visible when you upgrade.
A folder in the open Agent Skills format (agentskills.io): a SKILL.md with instructions, optionally bundling scripts, references and assets. skillsvault also governs Claude Code subagents and slash commands. Every published version is immutable and stored by content hash.
Distribution works in the 17 agents that load Agent Skills — Claude Code, Codex, Cursor, Windsurf, Copilot, Gemini, Cline, Roo, Continue, OpenCode, Zed, Amp, Pi, Antigravity, Augment, OpenClaw and Hermes: `skillsvault pull --all` installs approved skills and removes banned ones. The runtime gate, which blocks a banned skill or subagent at the moment it is invoked, uses Claude Code's hooks today. Other agents get it as soon as they ship a blocking hook.
It covers six of the ten in full — malicious skills, over-privileged skills, insecure metadata, update drift, poor scanning and missing governance — and three in part. It doesn't sandbox skills (AST06), because it doesn't run them: your agent does. The full mapping, risk by risk, is at skillsvault.io/docs/owasp.
Only the audit upload. The decision is deterministic and computed on the machine against a signed policy bundle, with no model and no network call. The gate writes its decision, then records it in your audit log with one HTTPS request, capped at 1.5 seconds. If the control plane can't be reached, the decision stands and the event is queued for the next sync.
Skills, policies and the audit log live in your organisation's tenant, isolated by row-level security in an EU-hosted (Ireland) database. Gate decisions are made on the machine; each decision is then recorded in your audit log, exportable as tamper-evident CSV or JSON on Business and Enterprise. The AI features (the AI review at publish time and AI assist in the editor) send the skill's content, plus the author's request for an assist, to our model provider (OpenRouter or Vercel AI Gateway); your developers' agent prompts are never sent, and the gate and the audit log don't use a model.
Yes, on Team and above. A member proposes a skill or a change from the editor or with skillsvault publish. Owners, admins and governance officers review the diff, what the skill does and what the gate would decide, then approve it or send it back with a note. Nothing reaches a machine until it is approved.
The block links to a request form with the policy and skill filled in. A reviewer grants an exception for 7, 30 or 90 days, or declines it with a note. A granted exception exempts that one skill from that one policy on every machine, MCP and pull included, until it expires or is revoked. Every other policy still applies, and a skill banned in the registry stays blocked. Each step is recorded in the activity log.
Yes, through MCP. Add https://skillsvault.io/api/mcp in any MCP client that supports sign-in (OAuth). You sign in to skillsvault and pick your organisation; the client can then list and load the skills your organisation approved, read-only, with your policies applied and every load audited under your name. You can disconnect it from your account page at any time. Headless setups use an agent key instead.
Yes. Switching between Team and Business, or between monthly and yearly billing, on the billing page changes your existing subscription in place, and Stripe prorates the difference. To move to the free plan, cancel in Manage billing; your organisation moves to Free when the subscription ends, on the date Manage billing shows.
Govern every skill your agents run.
Set up your registry and your first policy in minutes. Free to start, no card required.