Build once. Reuse everywhere.
Skills are the atomic capabilities every AI employee draws from — search a wiki, open a ticket, provision an account, refund an invoice. One catalogue, one identity model, one audit trail. 48+ shipped, and one open SDK to write your own.
Six kinds of skill.
Every skill in the catalogue falls into one of six intents. Read down the group titles and you get the platform's working vocabulary in a glance.
Grounded lookups against your corpus, your warehouse, and your systems of record.
Deliver the answer, the update, or the request — through whatever channel the reader lives in.
The write-side skills that turn a request into a state change — always through the approval matrix.
Structured writes into your systems of record — CRM, ITSM, docs, code.
Reasoning-heavy skills that read many sources and produce one output.
Operational skills for the reviewer, not the requester — approvals, escalations, incident coordination.
One skill. Every agent that needs it.
A skill is not agent-specific. Ivy calls `provision-okta-user` for a new hire; Sig calls the same skill during a step-up review; a custom automation agent calls it on a schedule. Guardrails, permissions, and evaluation travel with the skill — not with each caller.
- One implementation · many callers · one audit trail
- Guardrails and PII rules live with the skill
- Skills call other skills — no reinvention
- Versioning + deprecation windows on the contract
Skills are how your platform team makes the AI workforce yours.
Every organisation has one or two capabilities no vendor ships — a bespoke ticket-triage rule, an ERP-specific write, a compliance skill the auditors know by name. The SDK is designed so those skills feel like first-class citizens beside the ones we ship, not bolt-ons around them.
See the developer platform
Eight parts. One contract.
Every skill in the catalogue is composed of the same eight primitives. Change any one and the skill's behaviour changes — versioned, evaluated, and rolled forward through the same control plane.
The system prompt that shapes voice, refusal behaviour, and when to escalate.
The APIs, MCP servers, or database queries this skill can call.
The corpora this skill reads from when the answer needs grounding.
Typed inputs — email, ticket ID, entity path. Identity fields are injected, not model-supplied.
Typed outputs — the shape callers depend on. Versioned; breaking changes are opt-in.
Scoped access — a skill inherits the employee's ceiling and can narrow further.
Input and output filters — PII masking, prompt-injection screens, banned-content checks.
Golden-set evaluation that runs before promotion. Accuracy, tool-success, hallucination rate.
The guarantees that make a skill safe to reuse
Write your own. In an afternoon.
The SDK ships the same primitives the built-in skills use — typed inputs, typed outputs, permission scoping, guardrails, evaluation. Write a skill in TypeScript, ship it via the CLI, roll it out by percent or cohort. Or expose your internal tool as an MCP server and every agent in your workspace can call it.
- Typed schemas — no untyped tool calls
- Test locally with `bq skill run`
- Golden-set evals gate promotion
- MCP bridge — expose any tool to any agent
Every skill carries a risk tier.
Read-only skills default to autopilot. Reversible writes route through the approval matrix. Destructive writes add step-up MFA on the reviewer before execution. The tier is a property of the skill, not the caller — Ivy or a custom agent, same gate.
Search, fetch, summarise — auto-run.
Update fields, create tickets, add-to-group — routed for approval.
Revoke, delete, transfer, refund — step-up MFA on the reviewer.
Skills your workforce actually runs on.
Book a demo and browse the catalogue on your own systems. First custom skill live in an afternoon.