One terminal, to rule them all...
- Many agents, one window. Run Claude Code, Codex CLI, Gemini CLI, Copilot CLI, Cursor Agent, opencode, Aider, Qwen Code and Goose side by side, in tabs, split panes and tab groups you can fold and drag.
- Paw Agent. A built-in agent on your own models: Anthropic, GitHub Copilot, ChatGPT, OpenRouter, Ollama or any OpenAI-compatible provider.
- Plans. Describe a goal with
/plan; a planner splits it into steps for your agents, you review the plan, and it runs them in order or side by side, shown as a graph (or in 3D). Steps that need you wait for your approval. - Factories. Pipelines of agents you describe once in a YAML file, started by a schedule, a watched command, a webhook or a file change. See Factories below.
- A shell in every tab. Run commands next to the agent in your own shell, then pass their output to the agent with one click.
- Keeps running. Agents run in a background process with a menu bar icon, so closing the window doesn't stop their work.
- Sessions. Each session's conversation, scripts and files are kept in
~/.paw; reopen past sessions and see your usage per day, week and month. - Plugins. Extend Paw with commands, status bar items and plan board views, from the official catalog or your own.
- Sandbox. Paw Agent can work in macOS's sandbox: changes only in its project folder, only the programs you allow, and the network only if you let it. See Sandbox.
- Your look and language. Themes for light and dark, your choice of font and size, and Paw in English, German or Turkish; Paw Agent answers in the language you pick.
- Your keys. Change Paw's shortcuts in Settings › Keys: click one and press the new keys. A key that's already taken (by another shortcut or by macOS) is turned down, with what uses it.
A factory is a pipeline of agents that Paw runs in the background, so it keeps going while the window is closed. You describe it once in ~/.paw/factories/<name>.yaml, or ask Paw Agent to write it:
name: Create Release
folder: ~/Developer/my-app
trigger:
watch: ./new-commit.sh # a run for each new commit
every: 30s
fires: each
nodes:
- id: bump
title: Bump the version and update the changelog
prompt: Bump the minor version, add a changelog entry for the new commit, and commit it.
- id: publish
title: Push to origin
needs: [bump]
tools: [shell]
prompt: Push the current branch to origin.- Any shape: nodes run in order or side by side; they can join (
join: allorany), branch (whenan earlier answer matches) and wait (waituntil a command succeeds). - Each node its own agent: its model, skills, built-in tools and MCP servers.
- Started on their own:
every,at(times of day),watch(a command),webhook(POST 127.0.0.1:4747/factories/<id>, its body going to the nodes) andfile.repeatloops a pipeline. - Side by side: every run has its own execution id (a UUID) and folder,
~/.paw/factories/<id>/<execution-id>/. Runs of one factory can go at once, each in its own project folder and with its own task. - Started from a chat: ask Paw Agent to run a factory on the project at hand; it starts the run once you confirm, and tells you how it's going.
- They ask you: a node can ask a question or wait for you to approve its answer, with a notification; answer in the view or in the node's own session.
- Watched like a plan: the graph shows each node's state (failed ones glow red, waiting ones amber), a click shows a node's task and activity, and the timeline shows where the time went. Approvals wait in the view until you answer them.
The full format is in docs/factories.md.