Nob is an agent platform built on Claude. Type a request in plain words and walk away: agents plan it, build it on your machine and check it against your own build and tests, without stopping to ask. You get one result back — what was done, how it was checked, what was assumed — and the work waits on its own branch for one Merge.
Dark mode toggle
Pick a spec and dispatch it. The console fills with the records a real task produces.
One line in the dashboard is enough. Nob reads your code, sizes the request, and plans a big one as several tasks on its own.
Nobody is asked anything along the way. When an agent hits a question it takes its best reading, writes down the assumption, and keeps going.
A driver agent gets a brief, works in a fresh worktree, runs your gate command, and posts checkpoints; you can steer it mid-run. The server records claims; it never touches your files.
One email when it is done: what it did, the checks it ran, the assumptions it made and anything it could not check. The work waits on its own branch for one Merge.
Add the Nob MCP server to your Claude Code config with an API key from the dashboard. Every dispatched task runs in a spawned subagent, never in the conversation you are typing in.
/nob Add a dark mode toggle
## Acceptance Criteria
1. ...Claude Desktop can connect too, but it has no way to spawn a subagent, so it can read and plan without dispatching.
Settings → API keys → Connect a machine hands you a one-time command. Paste it on the machine that holds your code and it pairs, registers your checkouts as projects, installs a background service, and starts polling.
curl -fsSL https://github.com/n0b0dd/nob-worker/releases/latest/download/install.sh \ | sh -s -- connect <PAIRING-CODE> --server https://api.ahhnob.com/mcp
Carries its own Node and Claude runtime, needs no database and no inbound ports, only writes inside the roots you list, and updates itself between tasks. On Windows the same panel hands you a PowerShell line instead (beta; needs Git for Windows). nob-worker on GitHub.
Every request comes back with the real output of your build and test commands, an independent check of whether it does what you asked, and the assumptions it made while nobody was watching.
A request too big for one agent becomes a plan of dependent tasks that run one after another on a single branch, re-planning when a task fails. You merge the whole thing once.
Work happens on its own branch in its own worktree. In Auto, passing work within your scrutiny band lands on its own and anything heavier waits for you. Want a say along the way? Work in HITL.
Pair a machine with one command. It picks up what you hand off without a Claude Code session, survives reboots, updates itself, and rotates between Claude accounts as they cool down.
Token usage and dollars are recorded per task, main run and subagents alike, with a usage meter that mirrors your Claude plan limits.
The whole product is in the repository — there is no private tree and no license check. Paying for hosting buys operations, not features.
Free
Every feature, no limits
Solo and Team plans
Same code, we run it