LAC — Local Agent CLI

A terminal coding agent that runs against the model you already have: a ChatGPT subscription, a local OpenAI-compatible server (llama.cpp, vLLM, LM Studio, Ollama), or Ollama's free cloud models. Fork of Gemini CLI (see NOTICE).

Install

npm install -g @yoonseoklee/lac
lac --version

Requires Node.js 20.5 or newer (node --version; the current LTS from https://nodejs.org is the simplest choice). An older Node.js is refused at startup with a message that says so. Nothing else to set up.

First run

lac

Inside the CLI:

| Command | What it does | |----------|---------------------------------------------------------------------| | /auth | Choose the auth method (local server / Gemini API key / Vertex). | | /model | Pick a backend. Ships with three entries: | | | codex — ChatGPT subscription | | | Local-sLM — your own OpenAI-compatible server | | | Ollama-Cloud — your local Ollama daemon: every model it serves, | | | local weights and ollama.com cloud models alike |

Start on a specific backend: lac -m Ollama-Cloud.

Picking a server in /model opens the list of models it serves right now. Ollama cloud models run under the account the daemon is signed in with: on a machine that has not signed in, /model shows the server as not signed in (cloud models refused) and offers Sign in to ollama.com (or run ollama signin).

Connect your own model

Nothing personal ships in the package. Every endpoint, model name and key is read from a .env file. The first run creates it for you at ~/.localagent/.env (Windows: C:/Users/<you>/.localagent/.env) from the packaged template and prints that path; open it, fill in the block for the way you connect, and run lac again. It applies from any directory; a .env in the project folder is also picked up and wins.

# Local / on-prem server (vLLM, Ollama, llama.cpp, LM Studio ...)
# URL must end with /v1. Leave the key empty for a keyless server.
LOCAL_sLM_BASE_URL=http://your-server:8000/v1
LOCAL_sLM_API_KEY=
# the model is chosen in /model

# Any OpenAI-compatible cloud endpoint (adds a "Cloud LLM" entry to /model)
CLOUD_LLM_BASE_URL=
CLOUD_LLM_MODEL=auto
CLOUD_LLM_API_KEY=

# Official APIs by key only
OPENAI_API_KEY=
GEMINI_API_KEY=
ANTHROPIC_API_KEY=

# More OpenAI-compatible endpoints: name them. Each appears in /model as "Cloud <NAME>".
CLOUD_LLM_TEAM_BASE_URL=
CLOUD_LLM_TEAM_MODEL=auto
CLOUD_LLM_TEAM_API_KEY=

The first run creates both of the files you edit, in one folder, and prints where they are:

~/.localagent/.env                 your keys and addresses
~/.localagent/server_config.json   model entries and MCP servers

Fill in only the block for the way you connect. The ChatGPT subscription needs no key at all: /model → 1. Subscription signs you in through the browser.

Those two are yours and are never rewritten. The originals they are copied from, root/.env.example and root/server_config.example.json, live inside the installed package, which every npm update replaces - which is why the copies are made outside it. A run after an upgrade that ships a changed original says so and leaves your file alone.

A project can carry its own server_config.json, which takes precedence over the one in the home folder, or an explicit path can override both:

LOCAL_AGENT_SERVER_CONFIG=C:/path/to/your/server_config.json

Per-entry request parameters (temperature, top_p, max output tokens, reasoning effort, extra body) live in one params block that is the same for every service type; what an endpoint refuses is learned from its own errors, never guessed from the model's name. Reasoning effort can also be picked in /model.

An explicit environment variable always wins over the packaged file. Sessions and tokens live under your home directory, never inside node_modules, so npm update keeps them.

License

Apache-2.0. Upstream copyright Google LLC (Gemini CLI); modifications copyright YSLee.