You watch it work instead of guessing
Every command runs in a terminal card you can see, every page opens in a browser card, and the plan sits on a to-do card that ticks itself off. Nothing scrolls past unseen inside the agent’s own window.
Canvas mode
Give an agent a task.It builds its own workspace.
Switch an agent card to canvas mode and it stops working in a hidden shell. It opens the cards it needs — a to-do list, a terminal for each job, a browser, helper agents — wires them with arrows and does the work where you can see every step.
- One click in the card menu
- All 19 agent CLIs
- Holds in dangerous mode
A short story in the NeuroSquad app. You add a Claude Code card, turn on canvas mode in its card menu and type a task. The agent then builds its own workspace: a to-do card for the plan, a terminal for the dev server, a browser card to check the page, and a Codex helper with a terminal of its own for the tests — each connected by an arrow it draws itself. When it finishes, all of those cards are still on the canvas.
8. Everything stays Done. The plan, the server, the page and the tests are all still on the canvas — click any card to see what happened there.
Why work this way
You see the work, not just the result
In canvas mode, everything the agent does outside its own files happens in cards on your canvas. That changes what you can see, what you can stop, and what the agent is allowed to touch.
Every tool is a card you can take over
The terminals are real shells in your project folder, the browser is a real browser. Type into them, stop the dev server, click around the page yourself — the cards the agent made are yours to use.
Arrows are permissions
An agent uses only the cards wired to it, and a helper sees only the cards wired to its own card. The agent may connect cards it or other agents created; terminals, browsers, MCP servers and Telegram bots you set up yourself stay out of its reach until you connect them to it.
Its own shell and web are switched off
Canvas mode doesn’t just ask nicely. For every supported CLI, NeuroSquad removes or denies the agent’s built-in shell and web tools — Claude Code gets Bash, WebSearch and WebFetch denied at launch, Codex loses its shell and web search, and Cursor CLI’s hook blocks them live, on every call. Dangerous mode skips approval prompts; it doesn’t bring them back.
Helpers in parallel, each with its own context
When the work splits, it starts helper agents — any agent CLI installed on your computer, Codex or Gemini next to Claude — hands each a task and collects the answer along the arrow. Every helper is its own session with its own context window.
Nothing gets lost
The cards stay on the canvas after the agent is done: the plan with its ticks, the server’s output, the page, the test run. The to-do and note cards outlive the agent’s own context being compacted — and you can rewire, restart or remove any of them.
Works with
Every agent CLI NeuroSquad runs
Canvas mode is in the card menu of all 19. Each CLI loses its shell and web in its own way — a launch flag, a config layer, a plugin or a hook — and NeuroSquad uses the one that holds, dangerous mode included.
Plain terminal cards have no canvas mode: there is no agent in them to restrict.
Good to know
Is canvas mode a sandbox?
No. It decides how the agent reaches the shell and the web: through cards you can see. The agent still reads and edits your project’s files with its own tools, and the commands it runs in a terminal card are real commands on your computer.
When does it take effect?
The card tools and the instructions switch at once, while the agent is running. The CLI’s own shell and web tools are taken away from its next launch — except in Cursor CLI, where a hook checks the card on every call, so they are blocked right away.
Does it still hold in dangerous mode?
Yes. Dangerous mode skips the agent’s approval prompts; it doesn’t give back the shell and web tools canvas mode took away. And an agent in dangerous mode can’t wire itself to an MCP server card on its own — only you can draw that arrow. One caveat: Goose keeps its extension manager, so the model could still turn on another extension.
How many cards can it create?
Up to 24 per agent in canvas mode and 48 per workspace — a guard against a runaway agent. Past that it has to ask you. It is told to reuse the cards it already has rather than open a second one for the same job.
How do I turn it off?
From the same card menu: Disable canvas mode. The cards it created stay where they are.
Let your next agent set up its own workspace
Free for Windows and macOS. Add an agent card, open its menu, turn on canvas mode.
Windows 10 / 11 · macOS 12 or newer, Apple silicon and Intel