Skip to main content
A Tool block is a capability the Agent can pick during a run. Tools come in two runtime kinds:
  • HTTP — call an external API.
  • Inline — run a snippet in JavaScript, Python, or SQL, sandboxed on our runtime.

The four things every Tool needs

  1. Name — how the Agent refers to it. Short and verb-y (refund_charge, lookup_order).
  2. Description — a one-liner the Agent reads when deciding whether to call it. This is the most important field. The Agent won’t call a tool it doesn’t understand.
  3. Parameters — typed inputs, each with a name, type, required/optional, and their own descriptions.
  4. Implementation — the actual call: URL + method for HTTP, or the code body for Inline.

HTTP tools

Configure:
  • MethodGET, POST, PATCH, PUT, DELETE.
  • URL — supports {param} templating for path segments.
  • Headers — including secrets (see below).
  • Body template — for POST/PATCH/PUT. Use {param} for interpolation.
Response is passed back to the Agent as JSON. If the API returns non-JSON, you can add a small post-processor in the Tool.

Inline tools

Write a function body in JavaScript, Python, or SQL. Parameters are auto-injected as named variables. The return value is what the Agent sees. Example (JavaScript):
Inline runs in a sandbox with a modest CPU + memory budget and no filesystem or network by default. Whitelist outbound hosts in the Tool’s settings if you need them.

Parameters

Each parameter has:
  • Typestring, number, boolean, enum, object, array.
  • Description — one line, read by the Agent when choosing values.
  • Required / optional — required params are always filled; optional are filled only when the Agent thinks they’re useful.
  • Default — used when the Agent omits an optional param.
  • Enum values — for enum type, list the allowed literal values.
Reorder parameters with drag-handles. Order affects presentation, not semantics.

Secrets and headers

Any header or body value can be marked secret. Secrets are:
  • Never shown in the UI after save (they render as ••••).
  • Never included in exported Flow JSON.
  • Injected at call time from encrypted workspace storage.
Rotate a secret from the Tool’s Inspector — old runs continue to work, new runs use the new value.

Approvals

Toggle Require approval before call for high-risk tools (money movement, destructive DB ops, notifications). Held calls appear in the Approvals queue with the exact arguments the Agent picked.

Testing a Tool

Every Tool has a Test panel. Fill in the parameters, hit Run, and see the raw response. Uses your Draft credentials, so you can iterate safely.