Palette agent
Last updated:
Palette's own agent, built into Palette Desktop. Instead of wrapping a third-party tool, it is a single harness you point at the model you want, and you choose where that model runs: on Palette Cloud, on your own machine, or on your own server. It comes with Palette, so there is no separate account, sign-in, or agent to install, and it is the default agent, so you have one ready from your first chat. Pick it per chat like any other agent.
Setup
The Palette agent is built into Palette Desktop, so unlike Claude Code or Codex there is no separate agent to install. Once Palette Desktop is installed, open a chat and choose Palette as the agent.
Then choose a model and where it runs, in Settings or from the chat picker:
- Palette Cloud needs no download. If you are signed in to Palette, you can start straight away.
- On-device needs a one-time model download of a few GB.
- Self-hosted needs your own model endpoint.
Models and where it runs
The Palette agent is a harness: you choose a model, and you choose where it runs. There are three ways to run it.
| Where it runs | Models today | What you need |
|---|---|---|
| Palette Cloud (experimental) | DeepSeek, with more on the way. Models Palette hosts for you. | Nothing to download. Runs on your Palette account, metered by usage. |
| On-device | Qwen (an open model), with more on the way. Runs locally on your machine. | A one-time model download of a few GB. Then it is free, private, and works offline. |
| Self-hosted | Whatever your own server offers. | Your own OpenAI-compatible endpoint, for example LM Studio or Ollama, plus an optional key. |
On-device is the most complete option today. Palette Cloud is live but still experimental, so expect it to change. More models are being added to both Palette Cloud and the on-device catalog.
Usage & billing
The Palette agent runs on Palette-provided models or your own compute, not a third-party subscription.
- Palette Cloud usage is metered on your Palette account.
- On-device and self-hosted run on your own machine or endpoint, so there is no per-token charge from Palette.
Limits & notes
- Palette Desktop does not currently guarantee that the Palette agent loads
AGENTS.mdautomatically. Mention the file in your prompt when the agent should use those workspace instructions. See Working with CLAUDE.md and AGENTS.md. - The Palette agent supports plan mode. Slash commands remain specific to Claude Code.
- An on-device model downloads once and runs locally, so the first message can take a few seconds while the model loads.
FAQ
Do I need to install or sign in to anything?
The Palette agent comes with Palette Desktop, so there is no separate agent to install like Claude Code or Codex. To use Palette Cloud you just need to be signed in to Palette. For on-device, you download a model once first.
Can I run it fully offline and privately?
Yes. Choose an on-device model. After the one-time download it runs locally on your machine, so it is private and works offline.
Can I use my own model server?
Yes. Choose self-hosted and point Palette at your own OpenAI-compatible endpoint, for example LM Studio or Ollama.
Which models can I use?
On Palette Cloud, DeepSeek today, with more coming. On-device, an open model like Qwen that you download once. Self-hosted, whatever your own endpoint serves.
Related
- Agents, all five agents and how model access works.
- Chats, pick an agent per chat.
- Quick Start, install and run your first session.