Network Boundary

Inference runs on your Mac. It needs no API key, model provider, or account. A few features use the network, and each one starts with something you asked for. This page lists every path.

What stays on your Mac

• Inference. The interface sends chat completions to http://127.0.0.1:PORT/v1, a llama-server process started by localcode from the binary included in the wheel. The provider is fixed to that server; there is no setting that points inference elsewhere.
• Your files, prompts and edits. The agent reads and writes your working tree directly.
• Sessions and logs. Sessions live in <project>/.localcode-agent/. Server and supervisor logs live in the run directory. Nothing is uploaded.

There is no analytics endpoint, usage reporting, update check, or account. Every localcode process binds to 127.0.0.1. Loopback alone is not a boundary, since a web page can send requests to a local port. So each session also has two random secrets: the model service refuses any request without the session token, and the model server refuses any request without the session key. Both are generated at launch and kept owner-only on disk.

Where localcode uses the network

#	What	Where it goes	When
1	Model download	huggingface.co	Only when you press Enter on a Download row in the picker. The size is shown first
2	Quant listing	Hugging Face repo API	When you open a model in the picker, to list the quants its repo ships
3	Image projector	huggingface.co	Only when you run /vision and confirm the size
4	Language server	The server’s own release source	Only when you install one through /lsp and confirm. OPENCODE_DISABLE_LSP_DOWNLOAD=1 is the default, so nothing is fetched otherwise
5	Voice	The recorder runtime (about 60 MB) and speech model (about 514 MB)	The first time you use voice input, after a consent screen that lists both
6	websearch tool	DuckDuckGo by default; Exa or Parallel if you set a key	Whenever the model calls it
7	webfetch tool	The URL named in the call	Whenever the model calls it. Page text is capped at 20,000 characters (LOCALCODE_WEBFETCH_MAX_CHARS)
8	MCP servers	Wherever you pointed them	Whenever the model calls one of their tools
9	Shell commands	Wherever the command goes	Whenever a bash call runs; commands inside the project do not prompt

Web search works out of the box with a keyless DuckDuckGo backend. To use a paid search API instead, set LOCALCODE_ENABLE_EXA=1 with EXA_API_KEY, or LOCALCODE_ENABLE_PARALLEL=1 with PARALLEL_API_KEY. The key is sent only to that provider.

The web tools, MCP tools and shell commands all run without a permission prompt; see Permissions.