Offline

localcode is local-first. After a model is on disk, the things you use most work with no connection. A few features still need the network, and each one is something you choose.

Works offline

• Model inference. The interface only ever talks to the local llama-server.
• Reading, editing and writing files.
• grep, glob, and shell commands that do not reach the network.
• The project’s own typecheck, test and lint commands, which the plugin runs after edits.
• Sessions, /undo, /diff, /review, and the model picker for models already on disk.
• Voice input and vision, once their downloads are on disk.

Needs the network

• Downloading a model. The picker still opens offline and lists what is on disk; the Available to download section is empty or fails to load.
• /vision, the first time, to fetch a model’s image projector.
• /lsp, to install a language server. Starting one already installed works offline.
• Voice, the first time, to fetch the recorder runtime and speech model.
• websearch and webfetch.
• Remote MCP servers, and any shell command that downloads something.

Preparing for an offline session

1. Start localcode once while connected. Pick the starred model in the picker and let the download finish.
2. If you want images or voice, run /vision and try voice once so their downloads complete.
3. Run /status to confirm the model is loaded.
4. Install any project dependencies first - a pip install or npm install fails offline like any other network command.

When offline, websearch and webfetch return errors. The agent uses the failed result as context and keeps going. See Network Boundary for every path that uses the network.