Developer Tool Providers

This page covers providers that integrate through local assistant tooling or CLI runtimes instead of a standard remote LLM API. They are useful when LibreFang agents need to reuse an existing coding or terminal-assistant stack.

Included Providers

• Claude Code
• GitHub Copilot
• ChatGPT (Session Token)
• Kimi Coding
• Aider
• Claude Code CLI
• Codex CLI
• Gemini CLI
• Qwen Code

CLI logins as first-class default providers

When you have not set a provider API key but do have a working CLI login (claude, codex, gemini, qwen available on PATH and authenticated), LibreFang will pick the CLI provider as the default model for new agents. The detection runs at startup and is rechecked when the daemon refreshes provider availability.

Detection priority (in order, first match wins):

1. claude-code — claude binary on PATH, claude auth status reports an active session.
2. codex-cli — codex binary on PATH, with OPENAI_API_KEY (which the Codex CLI itself needs) set.
3. gemini-cli — gemini binary on PATH, authenticated via gemini auth login.
4. qwen-code — qwen binary on PATH, authenticated.

This behaviour replaced the previous "fallback when API providers fail" heuristic — CLI providers are now treated like any other remote provider. To explicitly override the auto-pick, set [default_model] provider = "anthropic" (or whichever) in config.toml.

Claude Code

Notes: Claude Code is an Anthropic model variant optimized for agentic coding tasks. It uses the same API key and base URL as Anthropic but targets models tuned for long-horizon tool-use workflows.

GitHub Copilot

Setup:

1. Subscribe to GitHub Copilot
2. Create a Personal Access Token with copilot scope
3. export GITHUB_TOKEN="ghp_..."

Notes: The Copilot driver handles OAuth PKCE token exchange automatically — it obtains a short-lived Copilot API token from https://api.github.com/copilot_internal/v2/token and caches it with auto-refresh. The Copilot API uses OpenAI-compatible chat completions format. Tokens are refreshed 5 minutes before expiry.

ChatGPT (Session Token)

Setup:

1. Run librefang auth chatgpt (browser-based OAuth flow). The CLI captures the session token and caches it.
2. The token auto-refreshes 1 hour before expiry; no manual rotation needed.

Minimal config.toml:

[default_model]
provider = "chatgpt"
model = "gpt-4o"

Notes: Uses the proprietary Responses API rather than /v1/chat/completions. This lets agents reuse a paid ChatGPT subscription without a separate OpenAI API key. Tools and vision supported. Embeddings are not available via this driver — use OpenAI hosted instead.

Kimi Coding

Setup:

1. Sign up at the Kimi console and enable the Coding endpoint
2. Generate an API key under the Coding product
3. export KIMI_API_KEY="sk-..."

Minimal config.toml:

[default_model]
provider = "kimi-coding"
model = "kimi-coder"

Notes: Coding-specialized variant of Moonshot's stack, exposed in Anthropic Messages API format (not OpenAI-compatible). Uses the same internal driver path as Anthropic, including tool-use schema conversion and vision support. Different env var from MOONSHOT_API_KEY — Kimi Coding uses its own key.

Aider

Setup:

1. Install Aider: pip install aider-install && aider-install
2. Configure Aider's LLM provider via its own env vars (e.g. OPENAI_API_KEY)
3. No additional LibreFang configuration needed

Notes: CLI Provider — LibreFang spawns the aider binary as a subprocess in non-interactive mode (--message). Aider handles its own LLM provider authentication via standard environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.). Aider's --yes-always, --no-auto-commits, and --no-git flags are applied automatically. Use AIDER_CLI_PATH to override the binary path.

Claude Code CLI

Setup:

1. Install Claude Code: npm install -g @anthropic-ai/claude-code
2. Authenticate: claude auth login
3. No additional LibreFang configuration needed

Notes: CLI Provider — LibreFang spawns the claude binary as a subprocess in print mode (-p). The driver strips other providers' API keys from the subprocess environment to prevent leakage. Active subprocess PIDs are tracked and message timeouts (default 5 minutes) prevent hung processes from blocking agents. Vision input is supported via base64-encoded images.

Codex CLI

Setup:

1. Install Codex CLI: npm install -g @openai/codex
2. export OPENAI_API_KEY="sk-..."
3. No additional LibreFang configuration needed

Notes: CLI Provider — LibreFang spawns the codex binary as a subprocess in quiet mode (-q). The driver strips other providers' API keys from the subprocess environment (preserving only OPENAI_API_KEY and CODEX_* variables). This allows users with Codex CLI installed to use it as an LLM provider without additional configuration.

Gemini CLI

Setup:

1. Install Gemini CLI: npm install -g @google/gemini-cli
2. Authenticate: gemini auth login
3. No additional LibreFang configuration needed

Notes: CLI Provider — LibreFang spawns the gemini binary as a subprocess in print mode (-p). The driver preserves GEMINI_* and GOOGLE_* environment variables while stripping other providers' secrets. No separate API key is needed when using Google OAuth authentication.

Qwen Code

Setup:

1. Install Qwen Code: npm install -g @alibaba/qwen-code
2. Authenticate: qwen auth login
3. No additional LibreFang configuration needed

Notes: CLI Provider — LibreFang spawns the qwen binary as a subprocess in print mode (-p). The driver preserves QWEN_* environment variables while stripping other providers' secrets. Supports streaming JSON output from the Qwen Code CLI. No separate API key is needed when using Qwen OAuth authentication.