Circuitry Logo
Circuitry
Open Circuitry

Coding Agent

A chat that can see your canvas and act on it — with your choice of model or command-line assistant.

Circuitry's coding agent lives in a chat panel you open beside whatever you're working on. It sees the document in the other pane — a workflow, a design, a code file — and your current selection, so you can point at things and say "this." It can run as Circuitry's own agent or hand the wheel to a command-line assistant like Claude Code, Gemini CLI, or Codex CLI.

What a coding harness is

A model can only write text. A coding harness (also called an agent harness) is the loop wrapped around a model that turns its text into work: it gives the model tools (read and edit files, run commands, search), feeds it what it sees (your files, your errors), lets it work step by step until the job is done, and shows you what it changed for approval. Claude Code, OpenCode and Codex CLI are harnesses. So is Circuitry's built-in agent — the one in this panel.

You don't have to pick one forever. This panel runs Circuitry's own harness or any of the command-line assistants you have installed, and each harness keeps its strengths: Circuitry's agent knows your canvas and asks before it changes anything; Claude Code brings its own way of exploring a codebase. Switch in the header whenever you like.

Why use Circuitry's agent harness

  • It sees your canvas out of the box. Your open document and your selection are part of every turn: point at nodes, strokes, cells or lines of code and say "this." A command-line assistant gets the same ability automatically — when Circuitry sets one up for you (or you connect it through CLI mode), the Circuitry connector is configured into it, so it can read your selection, edit nodes and round-trip its edits through diff approval with no extra step.
  • It changes nothing without you. Code edits land as accept/reject diffs, bigger work as a plan you approve first.
  • It can run any model — and any other harness. The model picker covers Circuitry cloud models, your own API keys, OpenRouter, local models on this machine (Ollama, LM Studio, MLX), Apple Intelligence, and — like a model with a pulse — the command-line assistants themselves: pick Claude Code as the model behind the agent and its way of working drives the turn. A cluster adds the local models on your other computers, listed under each machine's name.
  • It can lead a team. Give it Agent Teams and it hands reading and research to your other models — even free or local ones — keeping only their reports in its own context.
  • It is yours to inspect. Every turn is a tool log: what it read, what it ran, what it changed.

The command-line assistants have strengths of their own, and Circuitry's harness is built to combine them rather than compete: run one as a harness in its own right (CLI mode below), or as the model behind Circuitry's agent, or as a member of a team.

Opening the panel

From a workflow, click the Agent Chat button (the speech-bubble icon) in the toolbar along the bottom of the canvas. The chat opens in a pane next to your work.

You can also pair a chat with any document by dragging it into a split-screen — the chat becomes an agent the moment it's side-by-side with a workflow, design, or code file. See Agent Chat in Split-Screen.

Run as many as you need side by side: ⌘T (Ctrl+T on Windows/Linux) starts another chat in the mode you're already in, so a second command-line assistant is one key away while the first keeps working. The Open chats list keeps track of them, and can sit in the side panel or dock against the chat itself — see the Chat Tour.

Two modes

A toggle in the chat header switches between the two:

  • Agent — Circuitry's built-in agent. It uses the model you pick, reads your open document and selection, and makes changes through plan approval (it shows you what it's about to do) and diff approval (code edits land as accept/reject hunks). This is the default.
  • CLI — runs a command-line coding assistant right inside the panel. Switch to it with the header toggle. CLI mode needs a project folder open and a connected Circuit.

In CLI mode a dropdown lets you pick which assistant to run:

  • Claude Code
  • Codex CLI
  • Antigravity CLI
  • Gemini CLI
  • Qwen Code
  • OpenCode
  • DeepSeek Harness (opens its own app rather than a terminal)

Each shows whether it's installed on the machine running the Circuit. Choose Set up AI… to add one. See Claude Code and other coding agents.

From an iPad or a phone

CLI mode is not desktop-only. Pair an iPad, iPhone or Android device with a Circuit on your own Mac, Windows or Linux machine and the assistant runs there — your repositories, your toolchain, your installed CLIs — while you read its output and approve its work on the device. The bottom bar turns into a key row when the keyboard is up, which is what makes Esc, Ctrl and the punctuation a phone keyboard hides reachable with a thumb. Set it up in Connect your iPad or phone to your computer, and use Home Circuit to keep that machine reachable when you are away from its network.

The conversation isn't tied to the device that opened it. Because the assistant runs on the computer, the same session appears on every device you sign in from, mirrored live — start it at your desk, follow it on your phone on the way out, and take over by typing. The Open chats list shows the conversations running on that computer, including an On your other devices section, so several can be watched at once. Closing a window doesn't end anything: a conversation stops when the last device deliberately closes it, or when you stop it from any of them.

Pasting images

Both modes take pictures — a screenshot of a bug, a design reference, a diagram:

  • Agent mode — attach an image to your message by pasting or dropping it onto the message box.
  • CLI mode — paste, drag-and-drop, or right-click an image straight into the terminal. Use Ctrl+V on macOS or Ctrl+Shift+V on Windows and Linux. See Terminals.

Speaking instead of typing

In agent mode you can dictate a message rather than type it — hold Space in the message box, speak, and let go. See Talking instead of typing for the full set of spoken commands. CLI mode is a terminal, so dictation doesn't apply there.

Connecting a command-line assistant to your canvas

A command-line assistant on its own can read and edit files. Connect it to Circuitry and it can also see your canvas — look up a node by name, read what you have selected, create and edit nodes, and round-trip its file edits through diff approval — the same abilities the built-in agent has.

There is nothing to set up by hand. When Circuitry installs an assistant for you (Set up AI… in the CLI picker, or the setup in Use Your AI Subscription), it configures the Circuitry connector — the Model Context Protocol server — into that assistant at the same time. It's also bundled with the Circuit download.

An assistant you installed yourself just needs the connector added to it once, and Circuitry offers to do that for you:

npx @circuitry-ide/mcp-server setup

Then ask, in plain English, to work with your canvas — "add a node that…", "read my selection and…".

This is the reverse of the MCP node, which brings an outside tool server into an agent. Here you're exposing Circuitry's tools to an outside assistant.

Models for the built-in agent

The Agent mode has its own model picker. The star on a row in the list marks that model as a favourite; the list shows favourites first, then the models you use most. Your options:

  • Circuitry cloud models — OpenAI, Anthropic, and Google models, run through Circuitry's server-side keys.
  • Command-line models — a connected assistant (Claude Code and friends) can back the agent too, kept alive across messages for fast turns.
  • Local models — run entirely on your own machine with Ollama, LM Studio, or MLX models on a Mac (Circuit sets those up for you). See Using Local AI Models.
  • Models on your other computers — add a computer to a cluster and its models are listed under its name.
  • OpenRouter — one key that reaches models from many providers. See Choose and add models.
  • On-device — Apple Intelligence on Mac, iPad and iPhone, with nothing leaving the device.
  • Your own custom models — add any compatible endpoint.

Choose which models appear in Settings → Models, then pick one per chat from the model dropdown. You can switch models mid-conversation — start cheap, bump to a stronger model for the hard part.

What you can ask it to do

Because the agent sees your canvas and selection, you can work by pointing rather than describing:

  • Look things up — "find the node called Fetch Orders" — the agent looks up nodes by name or searches for them.
  • Refer to a selection as "this" — select nodes, drawing strokes, spreadsheet cells, or lines of code, then ask about this or tell the agent to change this. Your selection is handed to the agent automatically.
  • Code → diagram — "draw a flowchart of the login flow in my code" — it reads the code and lays out a flow on a canvas.
  • Diagram → code — "implement this flow" — it reads the diagram and writes the code, landing as diff hunks.
  • Design → app — select design elements and ask the agent to integrate them into the app you're building.
  • Test in the browser — "open localhost:3000 and check the signup form works" — it opens a Browser tab, clicks through, and reads the browser console when something breaks. You watch its pointer move and can stop it at any point.
  • Everything in between — create and connect nodes, run code nodes, edit sheets, generate pages and designs, and search your knowledge base.

Related