Build
Tool anatomy
How one cURL command becomes a reviewed, versioned, governed AI tool.
Tools are what make your agent useful: real API calls it can perform. The import pipeline is built around one idea — a human decides what the AI may touch, once, and the platform enforces it on every call.
Import from cURL
Paste any cURL command into Anatomy. The dissection engine parses the method, URL, headers, and body, and classifies every part:
- Secret — API keys and tokens. Moved to the encrypted vault; injected server-side at execution; never visible to the model.
- Customer context — values that identify or describe the customer (IDs, emails, account numbers). Resolved per customer at runtime from the context vault — the model can neither read nor invent them.
- AI-fillable — the fields the model is allowed to fill (a search query, an order item, a quantity). Validated against the published contract on every call.
- Fixed — constants that are sent exactly as imported.
Review studio
The classification is a draft until a human confirms it. The Review studio shows every parameter with its classification, lets you reclassify (including in bulk), edit descriptions the model will read, and records every revision.
Validate and publish
Validation runs structural and security checks before publishing is allowed. Publishing creates an immutable version — the thing you bind to flows. Later edits create new versions; running flows keep the version they were bound to until you rebind.
Execution is fail-closed: a call that references an unresolvable secret, an unbound tool, or a field outside the contract is refused — not approximated.
Shape the response
Most APIs return far more than the AI needs — metadata, internal IDs, timestamps. In the Review studio, paste a sample response (or reuse one from a real run) and decide field by field what the model may see.
- Mark fields "Shown to AI" — once any field is shown, only shown fields reach the model (a strict allowlist, enforced server-side on every execution).
- For arrays, choose which item fields survive and cap the item count (e.g. first 50 menu items with just name and price).
- Add derived fields — count, sum, avg, min, max over an array — computed server-side and shown to the AI even when the underlying values stay hidden.
- A live preview shows exactly what the AI will see, with an estimated token saving.