An AI workspace for coding, research, and everyday work. Pix runs the Pi agent in a Docker Sandbox with your chosen models, reusable skills, and optional account connections.

Open a terminal in the folder you want to work on, run pix, and describe what you need. The agent can read and edit that folder, run tools, and delegate work to other models. Your files and saved conversations remain after you exit.

Pix's supported host setup is macOS with Homebrew, Docker Desktop, and Git. Start Docker Desktop, then install and sign into Docker Sandboxes:

brew install docker/tap/sbx
sbx login

brew install mcavage/tap/pix
pix setup

Setup guides you through choosing a model and connecting what it needs. You can rerun it if interrupted; completed setup is kept.

• API models: one OpenAI, Anthropic, or Google API key is enough. Store it in 1Password, install its CLI with brew install 1password-cli, and enable the desktop app's CLI integration. Paste the field's op://... reference when prompted, rather than the key itself.
• Ollama: if Ollama is running, setup offers supported installed models, including available cloud models. A local model needs no provider API key; Ollama Cloud uses your Ollama account.
• An environment supplied by your team: follow its setup instructions or add it below. It may provide models without personal API keys.

Start a session in an existing project or a new folder:

mkdir -p ~/pix-work
cd ~/pix-work
pix

Try a request such as “Summarize the documents in this folder” or “Review this project and suggest the smallest useful improvement.” Inside the session, /getting-started gives you a tour.

An environment supplies a set of models, skills, and connections. default is created for you; other names are entirely your choice.

For an environment you already have on disk:

pix env add ~/path/to/environment team
pix setup --env team
pix run --env team

pix env add also accepts a Git repository URL. Setup asks whether to make the new environment your default. Answer yes to use it whenever you type pix.

To update a Git-backed environment:

pix env upgrade team

This pulls the current branch forward and runs setup, preserving completed connections. Commit or stash local changes first. Restart your session afterward to use the updated environment. Linked local checkouts are updated in place; Pix does not pull updates automatically. You can change that choice later:

pix env default team
pix env list

Environment setup may open your browser to connect accounts. If it needs to run tools on your computer, Pix asks for your consent. Changes to that access can require approval again.

Ordinary sandboxes are removed when their last session exits. Your mounted project files and .pi-sessions conversations are kept; tools or files stored only inside the sandbox are discarded. Use --keep when you need those too. In scripts, use the explicit pix run command; bare pix requires a terminal.

Inside a session, /btw <question> asks a quick side question without interrupting the main task or adding the exchange to its context. The answer appears in a dismissible panel while the main agent keeps working.

Inside a session:

/remember Use concise summaries with the recommendation first.
/recall writing preferences
/forget <memory-id>

Memory persists across sessions. Capture is explicit by default. When Ollama's embedding model is available, recall can search by meaning as well as keywords; otherwise keyword recall still works. Setup can offer to prepare that embedding model. See memory for capture settings, scopes, and backups.

GitHub access is optional. To let the agent push branches and open pull requests, store a suitable token in 1Password and give Pix its reference:

pix secret set GITHUB_TOKEN op://vault/item/field

Replace the example with your token's actual reference. The next launch supplies it to that sandbox. The sandbox already includes git and gh; installing or logging into gh on your host does not configure Pix's sandbox credentials.

pix doctor

Doctor checks readiness and suggests fixes without changing your setup. Rerun pix setup to finish setup, or pix setup --env team to reconnect that environment's accounts. Add --verbose to setup, run, or doctor for diagnostics. If the problem is Docker Sandboxes itself, start with sbx diagnose.

Upgrade with brew upgrade mcavage/tap/pix. Your next launch updates Pix's own runtime and memory service as needed. Connection changes remain part of setup.

The agent can change files in mounted folders and use whatever account access you authorize. Sandboxing does not undo edits or prevent an authorized tool from acting on an external service. Review changes before committing or sharing them. Connected content and recalled memory may be sent to your selected model provider; see the security boundary.

Pix keeps its settings, environment definitions, credential references, and memory under ~/.pix. Set PIX_HOME to use a separate installation; cleanup is scoped to that home. Provider keys stay in 1Password, and Pix supplies them through sandbox-scoped credentials rather than host-global sbx secrets.

• First-session guide
• Command reference
• Memory
• Documentation index
• Contributing and maintainer instructions

MIT. See LICENSE.

Use plan to agree on the PR/FAQ, PRD and architecture before building a new product or material feature. build and deliver reuse that agreement; a brainstorm does not authorize production code. Existing approvals persist, and explicitly delegated product decisions can proceed autonomously.

One main agent owns the product and can implement directly. Ready independent units can run in parallel workers, followed by verification and independent review of the integrated product. Visual review uses actual screenshots; completion claims must point to checks that ran. Model choices remain in your environment or explicit overrides. See the product agreement rules.

See the delivery design for acceptance and local evidence collection.