Core concepts

Catalyst has a small number of moving parts. Once these click, the rest of the docs are just detail. Read this once and skim back to it when a term shows up later.

Models

A model is the AI doing the thinking — Claude, GPT, Gemini, or one you host yourself. Catalyst treats models as interchangeable: you pick one per conversation and can switch at any time. Some models are provided by your workspace (no key required); you can also bring your own. See Models & providers and Adding a provider.

The deliberate consequence: nothing you build is tied to a model. Your prompts, tools, data connections, and workflows all sit a layer above, so a better model next quarter is a setting change, not a rebuild.

Context

Context is everything the model can see when it answers: your current message plus the earlier turns of the conversation. It’s how follow-up questions work without repeating yourself.

Context isn’t unlimited — each model has a maximum size. When a conversation grows very long, start a new chat for a new topic. Catalyst also compresses long conversations automatically so you rarely hit the ceiling, but a fresh chat per topic keeps answers sharp.

Tools and agentic mode

By default a model only produces text. Tools let it do things — search the web, run code, generate images, query a database, or call an integration you’ve connected.

Agentic mode is what turns tools loose: instead of replying once, the model calls a tool, reads the result, decides what to do next, and repeats until the task is finished. That loop is the difference between “here’s how you’d do it” and “here, I did it.” You control which tools are available and how hard the model is allowed to think.

Artifacts

When the model produces something substantial — code, a document, a diagram, a working HTML page — it comes back as an artifact in a side panel you can preview, edit, and download. Artifacts are versioned: as the model revises, you can step back through earlier takes. This keeps real deliverables out of the chat scroll and in a place you can actually use them.

Workflows

A workflow is a chat made repeatable. Once you’ve worked something out by hand, you capture the steps — model calls, code, tools, conditions, loops — as a visual graph that runs the same way every time, on demand or on a schedule, and can deliver results by email.

The connection runs both directions: you can build a workflow from scratch, or promote a chat into a workflow once a conversation proves a process worth automating.

Connections

Connections are how Catalyst reaches your real world:

• Integrations (MCP) — connect external tools and services (your APIs, databases, knowledge bases) so the model can use them as actions. See Integrations.
• Models & providers — use the built-in models or bring your own with an API key. See Models & providers.
• Network devices — reach a model or service running on your own machine. See Network devices.

Your data also comes in more directly: attach files to a chat or a workflow, and let the model run code on them in a private sandbox.

Memory, skills, and prompts

These shape how the model behaves and what it knows about you, everywhere it runs:

• Memory — a profile plus a searchable store of facts and notes, so Catalyst recalls what matters across sessions instead of starting cold each time.
• Skills — reusable instructions the model can load on demand to handle a specific kind of task well.
• Prompts — your standing system instructions that set tone, defaults, and house rules.

Putting it together

A typical path through the product:

1. Chat to explore a problem, switching models as you like.
2. Turn on agentic mode and connect tools / data when you want the model to act on real systems.
3. Capture the result as a workflow when it’s worth repeating, and schedule it.
4. Let memory, skills, and prompts carry your context and standards across all of it.

Next: pick the surface you want to go deep on from the sidebar, or start with Chatting.