Components
The groups in the palette, what belongs in each, and how to find one when you do not know its name.
The Components rail is the palette. It is grouped by what a component does rather than by which vendor supplies it, so the same group holds several providers' versions of the same idea.
| Group | Holds |
|---|---|
| Input & Output | Chat input, chat output, text input — where data enters and leaves |
| Data Sources | APIs, databases, files, the web |
| Models & Agents | Language models, and agents that use tools |
| LLM Operations | Prompts, structured output, parsing a reply |
| Files & Knowledge | Loading, splitting, embedding, retrieving |
| Processing | Transforming data between steps |
| Flow Control | Branching, looping, conditions |
| Utilities | The small pieces — dates, maths, formatting |
| Prototypes | Experimental components that may change or disappear between releases |
A Saved group appears above these once you have saved a component of your own.
Two entries sit below the groups:
- Discover more components — the wider library, beyond what is installed.
- New Custom Component — write your own in Python when nothing fits.
Finding one
The search box above the rail matches on name and description, which is
usually faster than remembering which group something lives in. If you
know what the step should do but not what it is called, search the
verb: split, embed, parse, branch.
Connections
Components connect output to input, and the types have to agree. If a connection will not attach, the two sides are not compatible — the fix is usually a Processing component between them, not a different model.
Tools
An agent can be given tools: components attached to its Tools
input rather than wired in sequence. The Simple Agent template ships with
two — Web Search (PERFORM_SEARCH) and URL (FETCH_CONTENT) —
and neither needs a key of its own.
The agent decides when to call a tool. You are not wiring a path through them; you are handing over a set of capabilities and letting the model choose.
MCP and Bundles
MCP exposes servers implementing the Model Context Protocol as components, so an external service becomes something a flow can use. Nothing is registered by default; add a server there and its tools appear as components.
Bundles groups components by provider, for when you do know the vendor you want.