For Claude Code, Codex, OpenCode, Qwen, Cursor

Your AI agents,on one canvas.

NeuroSquad is a desktop command centre for CLI coding agents. Every agent is a live terminal card. Draw an arrow to a browser, a terminal, a note or a todo list — and the agent can use it. Pick a scenario and drive it.

Claude Code drives a real browser card through your webmail, drafts replies, and waits for your OK before sending.

10CLI agents supported
12+card types on the canvas
0accounts or cloud required

Works with the agents you already use

Claude Code
Codex
OpenCode
Qwen Code
Cursor
Kilo Code
Hermes
pi
omp
Crush
Claude Code
Codex
OpenCode
Qwen Code
Cursor
Kilo Code
Hermes
pi
omp
Crush
Arrows are tools

Cards that talk to each other

Draw an arrow from an agent to any card and the agent gets that card’s tools through NeuroSquad’s built-in MCP server. No config files, no plugins — the canvas is the wiring diagram.

Tap a card to see what the agent can do with it

Browser

A real Chrome the agent navigates, clicks, fills and screenshots.

browser_navigatebrowser_clickbrowser_fillbrowser_consolebrowser_screenshot
The browser card

A real browser your agent can drive

Not a screenshot tool — an actual Chrome window on your canvas. The agent navigates, fills forms, reads the console and network, takes screenshots and tries again after the fix. You watch every click.

Real Chrome, real sessions
Log in once; the agent works inside your session. Automation flags are off, so sites behave normally.
Console and network
Errors and failed requests come back to the agent as text it can act on.
Verify, don’t guess
The loop is reproduce → fix → click through again, visibly, on the same card.

The agent fills in a form in a real Chrome card, reads the console, takes a screenshot, fixes the bug and tries again.

Squads

One lead, many helpers

Arrows between agents are directed: the lead delegates, helpers report back. Mix harnesses — Claude Code leading Codex, Qwen and OpenCode — and give every task on the board an owner.

One lead agent splits the work, assigns each task on the board to a helper — Codex, Qwen, OpenCode — hands it over an arrow, and collects the results.

Delegate over arrows

agent_send and agent_read turn any agent into a helper of another.

Tasks with an owner

On the task board each task is assigned to one agent; kanban_write refuses to wipe other agents’ tasks.

Spinner on every card

The sidebar shows who is working, who is waiting for you and who is done.

Dev servers

Start it, see it, fix it — on one canvas

The Dev servers card lists what is listening on your machine. One click opens any of them in a browser card next to the agent, so it checks its own work in the running app — not in its imagination.

Every local server, live
Vite, Next, Django, a database — with the process behind each port. Refreshes on its own.
One click to a browser card
Open puts the page on the canvas; draw an arrow and the agent drives it.
Agents can ask too
With an arrow, `ports_list` tells the agent which port its server really came up on.

The agent starts the dev server, the Dev servers card picks up the port, one click opens it in a browser card — and the agent checks its own page.

Worktrees

Every agent on its own branch. A second opinion in one click.

An isolated agent works in its own git worktree, prepared before its first prompt: your .env copied in, your setup command run. When it is done, “Add a reviewer” puts a different model on the same tree — told to read and not to touch.

checkout-fixagent/checkout-fix
This agent’s worktree
Branch
agent/checkout-fix
Path
…/worktrees/checkout-fix
Setup command
npm ci
Preparing the worktree
  1. copy .env
  2. copy .env.local
  3. npm ci
A worktree per agent
Parallel agents never edit the same checkout. Your main branch stays as you left it.
Ready before the first prompt
Listed files are copied, not symlinked, and the setup command finishes before the agent starts.
A second model, not a second author
The reviewer gets an arrow from the author and a prompt that says: do not edit anything.
Flow timer & Telegram

Focus. The squad can wait — unless it can’t.

Start a focus block on the Flow timer card and “finished” pings wait until the break, then arrive as one digest. “Needs input” always comes through — and with a Telegram card it reaches your phone, where you can answer the agent directly.

checkout-fix
refunds-api
docs
signup-qa
Flow timer
25:00Focus
Holding “finished” pings until the break
Squad bot
bot
Tap to reply from the phone

Pings held, not lost

Everything that finished during the block is waiting for you at the break, agent by agent.

Blocked agents still reach you

An agent that needs you is never held — a stuck agent is wasted time.

Your own Telegram bot

Messages to the bot reach the connected agents as prompts; they can answer with telegram_send. No server, no webhook.

Usage & budget

See the spend. Cap it.

NeuroSquad reads token usage from the agents’ own transcripts and charts it per agent and per hour. Set a budget per workspace — when it is hit, every agent gets a Ctrl+C (not a kill) and automated prompts stop. You typing by hand is never blocked.

Claude CodeCodexQwen
$7
Mon
$9
Tue
$6
Wed
$11
Thu
$5
Fri
Sattoday
Sun
Spent this week
$38.00
Budget$60
Claude Code
Codex
Qwen
All agents running

Per-agent analytics

Tokens and cost by agent, by hour, by workspace.

Budget brake

Hit the limit → agents paused, sessions kept, queues held.

Raise to resume

Lift the limit above what’s spent and work continues.

Remote

Your whole squad, in your pocket

Turn on remote access and scan a QR code: your phone opens the full app — live terminals, cards, notifications — over your local network or through a Cloudflare tunnel. Approve an agent from the sofa.

Scan to pair

A QR code with a one-time token. The token never leaves the link.

Full app, live

Terminals stream in real time; type, approve, switch workspaces.

Anywhere via tunnel

Works over Wi-Fi, or through a Cloudflare tunnel when you are out.

9:41
NeuroSquad
shop-frontend
checkout-fixWorking
signup-qaNeeds input
DocsFinished
signup-qa
browser_fill(form: "signup")
⎿ 2 fields
Update(src/api/signup.ts)
>
Ctrl+Shift+SpaceHold
checkout-fix
Update(src/cart/total.ts)
>
Dictation

Talk to your agents

Hold a hotkey, speak, let go. Speech is recognised on your machine — Parakeet or Whisper, GPU or CPU — and typed straight into the active agent. Nothing is sent anywhere.

Push-to-talk or toggle

Hold to talk, tap to toggle. Works in any app.

On-device models

Parakeet TDT 0.6B by default; Whisper large-v3-turbo on demand.

Straight into the agent

Text goes into the focused card, and to the clipboard.

Beyond code

Agents for the rest of the business

The same cards run inbox, sales, back-office and automation work. Each page below is a live, clickable scenario.

Local-first

Your code stays on your machine

NeuroSquad has no backend and no account. It orchestrates the CLIs you already have, with the keys you already use.

No cloud in the middle
Workspaces, notes and history are JSON files on your disk.
Your own agents
We launch the harnesses you installed; your subscriptions and keys stay theirs.
Nothing deleted behind you
Removing a card never touches your project folder.
Open about limits
Where something is not supported, the app says so in plain text.

Put your agents on the canvas

Free desktop app for Windows, macOS and Linux. Bring your own agents.

The first-run wizard installs an agent CLI, Git and the dictation model for you — or skips each step.

Questions