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.

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

git clone https://github.com/0xSero/harness-bridge && cd harness-bridge
bun install
cd packages/cli && bun link      # puts `harness-bridge` and `hbr` on your PATH

# 1. name an endpoint
harness-bridge providers add --name HomeLab --url http://host:8080/v1 --key sk-…

# 2. see what it actually serves, and pick
harness-bridge models
harness-bridge use deepseek-v4.1-flash

# 3. open a harness on it
harness-bridge run claude

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.

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

open -a "/Applications/Local AI.app"

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.

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

harness-bridge run --harness omp --exec --dir /Users/you/project

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.

harness-bridge terminal                 # what is installed, and what is in use
harness-bridge terminal ghostty         # warp, terminal, iterm, kitty, wezterm, …
harness-bridge terminal custom --command 'open -a WezTerm {dir}'

auto, the default, follows $TERM_PROGRAM: launch from Warp and it opens in Warp; from Ghostty, in Ghostty. Each terminal has its own mechanism — Ghostty, kitty, Alacritty and WezTerm take the command as argv; Terminal.app and iTerm are driven by AppleScript; Warp has no -e, so it is driven by a Tab Config (warp://tab_config/<name>), which opens a tab in the window that already has focus and only opens a window when none is open. That file is the only thing this tool writes into another app's directory.

{command}, {dir} and {name} substitute into a custom template.

harness-bridge dir ~/code          # remembered for every later session
harness-bridge run --dir /tmp omp  # or just this one

With no dir set, a session opens wherever the shell started. The panel has a folder picker and the web UI a text field for the same setting.

Claude Code and Codex skip their own confirmation prompts by default — --dangerously-skip-permissions and --dangerously-bypass-approvals-and-sandbox. That is convenient against a local endpoint and worth knowing about, so it is not hidden:

harness-bridge args claude                 # what it launches with
harness-bridge args claude -- --permission-mode acceptEdits
harness-bridge args claude --default       # back to the built-in flags
harness-bridge run --safe claude           # omit them for this one launch

A harness is only offered when the endpoint speaks its dialect. That is a correctness rule, not a preference: Responses-shaped traffic to a Chat endpoint fails on the first turn.

Which dialects an endpoint serves is discovered, not declared: providers scan sends a one-token request to /chat/completions, /messages and /responses and records what answers. Running it after a provider changes — or an engine gains a dialect — is what unblocks a harness that was greyed out.

The context window a harness is told about is the model's real one, read from the endpoint's /models, so nothing silently assumes 128k. Claude Code does not know locally served model names, so the bridge also sets CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT — it then uses the real window instead of warning and compacting early.

auto by default: nothing is sent, and the harness and engine keep their own settings.

harness-bridge providers reasoning high

harness-bridge run --harness crush     # not installed -> installed, then launched
harness-bridge install crush           # or on its own

Installs use the vendor's own package (npm, bun or pipx). Harnesses that ship as downloaded binaries — OMP, Grok, Hermes — are not guessed at: the tool prints the vendor's install hint instead. --no-install refuses to install anything.

providers                              list configured providers
providers add --name N --url U --key K [--apis chat,messages] [--reasoning L]
providers rm <id> | use <id> | reasoning <level> [provider] | show <id> [--json]
models [provider] [--fresh]            live models from the endpoint
use <model> [provider]                 select a model
harnesses [provider]                   installed, and what the endpoint can drive
install <harness>                      install one
run [model] [harness]                  open a session
     --print  --exec  --no-install  --safe  --provider <id>  --dir <path>  -- <flags…>
pin <model> | unpin <model> | pins     what the panel shows
dir [path]                             where sessions open
terminal [id|auto|custom]              which terminal sessions open in
args <harness> [-- flags | --default]  the flags a harness launches with
snapshot [--json] | status | config    the whole state, a summary, or the file path

models and snapshot cache the model list on disk for 30 seconds, so a UI opens instantly instead of waiting on the endpoint; --fresh refetches now. When the endpoint fails, the last list is served rather than an empty one — a momentary blip should not blank the panel.

bun packages/web/src/server.ts     # http://127.0.0.1:4141

Providers, the model catalogue, and launching — the same commands as the CLI, in a browser. Bound to loopback with no authentication: it is for your machine, or for a tailnet you trust.

packages/core/src/core.ts        providers, sessions, launch plans
packages/core/src/harnesses.ts   what each harness is, how it installs, its default flags
packages/core/src/terminals.ts   how a session is opened in each terminal
packages/core/src/models.ts      the model list and its disk cache
packages/core/src/snapshot.ts    the read model every UI renders
packages/cli/src/cli.ts          command wiring
packages/tray/*.swift            the menu bar app
packages/web/src/server.ts       Bun.serve and the single page

bun test

Unit tests cover dialect gating, the argv and environment built for each harness, secret handling and config round-tripping. test/cli.test.ts runs every verb in a throwaway config home — the test that catches the bugs refactors actually introduce.

End to end, a real completion was returned through the generated launch environment on:

Headless hosts are fine for everything except opening a window: terminal and snapshot work, and a launch reports the missing emulator instead of failing silently.

MIT