claude-launcher

Launch Claude Code with multiple backends (Anthropic, OpenRouter, Ollama, NVIDIA NIM, LM Studio, llama.cpp).

Features

• Multiple backends - Anthropic, OpenRouter, Ollama, NVIDIA NIM, LM Studio, llama.cpp
• OAuth login - authenticate with claude-launcher login
• Model picker - searchable model selection
• Exacto support - auto-uses :exacto variants for better tool calling
• Role models - configure different models for sonnet/opus/haiku tasks
• New model alerts - notifies when new models are available
• Local backends - run fully offline via Ollama, LM Studio, or llama.cpp

Install

npm install -g claude-launcher
# or
pnpm add -g claude-launcher
# or
yarn global add claude-launcher
# or
bun add -g claude-launcher

Requires Claude Code installed.

Usage

claude-launcher                    # launch with saved settings
claude-launcher login              # authenticate with OpenRouter
claude-launcher logout             # clear stored credentials
claude-launcher -b                 # pick backend and model
claude-launcher -a                 # use Anthropic backend
claude-launcher -o                 # use OpenRouter backend
claude-launcher -l                 # use Ollama backend (local)
claude-launcher -n                 # use NVIDIA NIM backend
claude-launcher -s                 # use LM Studio backend (local)
claude-launcher --llamacpp         # use llama.cpp backend (local)
claude-launcher -k                 # use Kimi backend (Kimi Code subscription)
claude-launcher -x                 # use Mixed backend (Anthropic native + any backend per role slot)
claude-launcher -- --resume        # pass args to claude

Backends

• Anthropic - standard Claude Code, no extra config
• OpenRouter - any model via OpenRouter; OAuth login or OPENROUTER_API_KEY
• Codex - run Claude Code on a ChatGPT subscription (reuses the codex CLI login)
• Grok - run Claude Code on an X/SuperGrok subscription (reuses the grok CLI login)
• Kimi - run Claude Code on a Kimi Code subscription (reuses the kimi CLI login)
• Mixed - any backend per role slot in one session: main/opus/sonnet/haiku each stay on the Claude Pro/Max subscription or route to Codex, Grok, Kimi, OpenRouter, Ollama, NIM, LM Studio, or llama.cpp. Slot models use a backend:model prefix and requests are routed per model, so /model codex:gpt-5.6-sol or /model ollama:qwen3 mid-session works.
• Ollama - local models, auto-filtered to tool-capable ones
• NVIDIA NIM - cloud (NVIDIA_API_KEY) or self-hosted endpoints
• LM Studio - local models via the LM Studio server (host must include /v1, e.g. http://localhost:1234/v1)
• llama.cpp - a local or LAN llama-server instance; setup prompts for the server URL (server root, e.g. http://localhost:8080 or http://my-box.local:8080)

NIM and LM Studio run through an in-process Anthropic-to-OpenAI translation proxy. Ollama and llama.cpp speak the Anthropic Messages API natively, so Claude Code talks to them directly.

First Run

1. Run claude-launcher -b
2. Select a backend
3. Provide credentials if the backend needs them
4. Pick a model
5. Optionally configure role models (sonnet/opus/haiku)

Configuration

Settings stored at ~/.config/claude-launcher/config.json:

• Backend preference
• Selected models (main, sonnet, opus, haiku)
• API key (if logged in via OAuth)

Environment Variables

• OPENROUTER_API_KEY - fallback if not logged in via OAuth
• NVIDIA_API_KEY - NIM cloud API key

License

MIT