Harness Bridge

Local AI

Point any coding harness — Claude Code, Codex, OpenCode, Pi, OMP, Crush, Copilot, Grok —
at any OpenAI-, Anthropic- or Responses-compatible inference endpoint. Pick a model from
what the endpoint actually serves, click a harness, and it opens already talking to that
model.

Nothing you own is edited. Your ~/.claude.json, ~/.codex/config.toml and
~/.config/opencode are untouched: the endpoint, key and model travel in the launch
environment, and the only setting the tool persists is its own selection.

Four packages, one core:

Everything else is a thin shell over the core; each shell issues the same commands you
would type.

Quick start

Requires Bun ≥ 1.1; the menu bar app needs macOS with Xcode command
line tools.

The key is written to ~/.config/harness-bridge/config.json, mode 0600. It is never
passed on a command line, never logged, and never committed. A session never contains it
either — see Sessions.

macOS menu bar

bash packages/tray/build.sh builds Local AI.app and a Local AI.dmg; install from the
DMG, or:

The panel shows the model in use, everything pinned beside it, the folder and terminal a
session opens in, and every installed harness. Settings holds the providers — select one
to edit it; the key is never displayed and a blank field keeps the stored one — and the
full model catalogue, with pin toggles.

Sessions are a command

A session is not a generated script. It is one command you can read:

The terminal is handed exactly that. It re-enters the CLI, which builds the launch
environment at run time from the 0600 config — so no key is ever written to a file,
and a session always reflects your current selection rather than one frozen when a
command was composed. Delete ~/.config/harness-bridge/ and the tool leaves no trace.

Choosing the…