LLM Step
Description
Section titled “Description”The LLM Step sends a prompt to a large language model. With an Anthropic LLM connection, the model gets scoped access to the project tables, dimensions, and documents you bind — Read, Write, or both per object — and writes its results directly into the Write-enabled bindings. A Write binding is the step’s output; there is no separate output format.
For a full walkthrough, see the LLM Step guide.
Configuration
Section titled “Configuration”- LLM Connection (required) — a connection of kind LLM (for example, Anthropic). Its Agent Access must be Read & write for any Write binding. Full also allows writes until it is removed on January 15, 2027, and a step that uses it logs a warning when it completes. At Read & write a step writes only to bindings with Write checked, and cannot run SQL that writes or create tables.
- Model (optional) — defaults to the connection’s default model. Any Claude model works with an Anthropic connection, older models such as Claude Sonnet 4.5 included.
- Prompt (required) — supports
{{tables.NAME}},{{dimensions.NAME}}, and{{documents.NAME}}references to bound objects. - Result schema (optional for Anthropic; required for other providers) — a JSON Schema (root
"type": "object") for a captured structured summary; with Anthropic, leave blank when the model’s work is its MCP writes. Non-Anthropic providers have no MCP access, so a schema is required. - Bindings — Tables, Dimensions, and Documents the model may access, each picked with a selector and granted Read and/or Write. Write tables receive inserted rows, write dimensions receive nodes, write documents receive uploaded files.
- Limits — Max output tokens (1,024 to 128,000) and Credential TTL (60 to 3,600 seconds).
Behavior
Section titled “Behavior”- The model call runs in its own job. A short-lived, scoped credential is minted for the run and revoked when it ends.
- Access is gated per object and per checkbox: the model reads only Read-bound objects and writes only Write-bound ones — never an object you didn’t bind. Writing requires an existing destination object; the step doesn’t create one.
- Each run makes a billable provider call and writes into your bound objects.
What’s Checked When You Save
Section titled “What’s Checked When You Save”- Result schema — when you provide one, it must be valid JSON, a JSON object, and declare
"type": "object"at its root. Otherwise saving is refused, for example with The result schema must declare “type”: “object” at its root. - Bindings and the connection — only an Anthropic connection gives the model access to bindings. Saving a step with bindings on another provider’s connection is refused, with a message naming the provider and asking you to Remove the bindings or choose an Anthropic connection. A binding with Write checked needs a connection whose Agent Access allows writes. On a Read-only connection, saving is refused with The selected LLM connection has Read-only Agent Access, so a binding with Write checked could not write. Set the connection to ‘Read & write’ or uncheck Write on the bindings. A step saved through the API with no connection is checked against the workspace’s default LLM connection, which it also runs on: the one an admin marked as the default, or else the most recently updated. Without an admin-marked default, editing another LLM connection makes that one the default, and can refuse the step’s next save.