Features / Terminal and your PCs

An agent in every terminal

labb reads and searches your code and runs commands with your approval, in any folder, on your own models.

labb in a terminal: asked why Docker rebuilds every dependency, it lists the project, reads the Dockerfile and pyproject.toml, and asks before an edit that installs the dependencies before copying the code.

You stay in charge

It asks before it acts

Permissions: Read-only commands run right away, and others ask: once, always in this folder, or no with a reason. Edits ask the same way, with a colored diff.
Secrets: Reading a file that may hold secrets, like .env or SSH keys, asks every time.
Modes: /plan is read-only and ends with a plan for you to approve. Shift-Tab goes through the modes: default, accept edits (commands still ask) and plan.
Questions: When something is open, it asks you one question with a few choices and its recommendation first.
Undo: /undo takes back the last message and its file changes, from snapshots kept apart from your project's .git.
Worth knowing: Commands aren't sandboxed, so review what you approve.

Longer jobs

Built for work that takes a while

Goals: /goal gives the conversation an objective. The agent keeps working, within a budget of model calls, until it has checked every part is done, and backs “done” with evidence.
Todo lists and notes: It writes a todo list for jobs of 3 steps or more, and saves what took effort to find out in .labb/notes.md.
Helpers: It can hand a search or investigation to a helper agent with read-only tools, so the conversation stays small.
Project instructions: It reads AGENTS.md (or CLAUDE.md) from your folder, and /init has it write one.
Background commands: Dev servers and watchers keep running while it works; /jobs lists them.
Your setup: MCP servers, hooks, custom slash commands, skills, the browser and your own engines.

At a glance

The commands you'll use most

`labb / labb -c`	Start a conversation in this folder, or continue the latest
`labb -p "question"`	One answer, for scripts (`--output-format json` too)
`git diff \| labb -p "review this"`	What's piped in goes with the question
`@file, ! command`	Attach a file; run a command yourself and share its output
`/plan, /undo`	Plan mode; take back a turn and its file changes
`/goal …`	Keep working until an objective is done
`/context`	What the model gets now, in tokens per part
`/model, /engine, /memory`	Switch models or engines; see your memories
`/skills, /<skill>`	List the skills, or use one on purpose
`/loop 5m …`	Repeat a prompt on an interval, as long as the session lasts
`labb run "Frost watch"`	Run a Studio workflow, showing its steps
`Esc, Esc Esc`	Stop the reply; edit your last message

Your PCs

Start work on a PC from the browser

Run labb connect on a PC and leave it running. The PC dials out to the server, so it works from behind a router. Then start a new chat, pick the PC and folder under Where, and say what to do. The PCs page lists what's connected.

labb can do it from a chat too, asking you first.

labb connect ~/code/project ~/code/other

Only yours: Nobody else sees your PCs, admins included.
Only those folders: The PC itself refuses any folder you didn't name.
It asks in the page: Commands and edits ask you, once each.
Only you, logged in: An API key can't start a run, so a key leaked from one PC can't run anything on the others.

With a server

Connected to your lab

Install it from your server and it uses the server's models, your documents and your memories. Sessions show up under Terminal in the web chat.

Without one

On its own

It also works alone, talking straight to Ollama, OpenAI or Anthropic. Document search needs a server; everything else works. Install it.

More features

Want it on your own machine?

Get labb