Agent Chat
Agent Chat is Circuitry's agent harness. It lets you work with many models from different providers, set up agent teams, write code, and edit your Circuitry documents.
- Many models, many providers. Use Circuitry's models, your own subscriptions and keys, coding agents such as Claude Code, or local models on your own computers, and switch between them in the same chat.
- Agent teams. Let the model in charge hand tasks to other models. See Agent teams.
- Coding. Work in your project: read and edit files, run commands, and review changes before you accept them.
- Your documents. Create and edit sheets, workflows, diagrams, designs, notes and web pages.
Every chat starts with Ask. Type what you want, and the agent starts on it. It sees the document you have open and what you have selected, so you can point at something and say "fix this".
Start with Circuitry's agent
Circuitry's own agent works straight away, with nothing to install. Open a chat, or press ⌘T for another one, and ask:
- "Make a budget sheet for a small cafe."
- "Draw a flowchart of how a customer signs up."
- "What does this function do?" with some code selected.
You can keep typing while it works. It reads your message at its next step.
Decide how much it can do alone
The mode is at the bottom of the chat. Press ⇧⇥ to change it, or type /mode.
| Mode | What the agent does |
|---|---|
| auto | Gets on with it. Code changes come to you as a diff to accept or reject |
| ask | Also asks before writing files or running commands |
| accept | Applies code changes without showing a diff |
| plan | Only reads, then shows you a plan to run. Type /plan, or /plan followed by your request |
Choose who does the work
The model picker is in the chat header. The agent can run on any of these, and you can switch mid-conversation:
| You have | Use | Read |
|---|---|---|
| Nothing set up yet | Circuitry's models | Choose and add models |
| A Claude, ChatGPT or Google subscription | Claude Code, Codex CLI, Antigravity CLI | Coding agents |
| Keys or accounts with model providers | OpenCode, an OpenRouter key, your own endpoint | OpenCode and DeepSeek Harness |
| OpenClaw, or want an agent with its own memory and channels | OpenClaw | OpenClaw |
| A capable computer | Ollama, LM Studio, or MLX models on a Mac | Using local AI models |
| A second computer with a good graphics card or lots of memory | Its models, on all your devices | Use your other computers |
| Several of these | A team that shares the work | Agent teams |
Handy commands
Type / in a chat to see them all.
| Command | Does |
|---|---|
/model | Pick the chat's model. In a coding agent's chat, pick the model that agent runs on |
/effort | How hard the model thinks |
/mode | What the agent may do unasked |
/plan | Plan before acting |
/agents | Choose a team |
/mcp | Turn this chat's MCP servers on or off |
/clear | In a coding agent's chat, start the conversation over |
Related
- The coding agent: pointing at things, images, voice, and what to ask for
- Agent chat in split-screen
- Testing in the browser