freebird

Multi-provider free-AI router. Sends one task to several configured free providers in parallel, shows each provider's streaming output in its own tmux pane (opened in a new macOS Terminal window), and hands Claude a single summary.json to synthesise into one digested answer.

Trigger rules (important)

• Do not reach for this skill on your own initiative. It costs time and network quota.
• Invoke it only when the user types one of:
  • /freebird-review — code review on a file / directory (code-specialist models)
  • /freebird-extract — vision extraction from an image file (vision models)
  • /freebird-ask — general-purpose question (flagship chat models)
  • /freebird-reason — hard thinking (reasoning-specialist models only: DeepSeek R1, Phi-4-reasoning, QwQ)
  • /freebird-quick — fast consensus on a simple question (small fast models only)
  • /freebird-setup — key wizard
  • /freebird-doctor — health check on configured providers
• If the user talks about "free models" conversationally without a slash command, suggest running one — don't run it yourself.

Task profiles — which model per provider

freebird doesn't just pick any model with the right capability; each task has a hand-curated provider → model map in src/freebird/tasks.py. This means:

• /freebird-review deliberately picks dedicated code models where available (Qwen3-Coder on NVIDIA, Codestral on Mistral, Qwen2.5-Coder on Cloudflare/HF) and reasoning models on providers that have no code specialist (DeepSeek-R1-Distill on Groq/Together, Phi-4-reasoning on GitHub Models).
• /freebird-extract picks vision-strongest per provider (Llama 3.2 90B Vision on NVIDIA, Pixtral Large on Mistral, Qwen2.5 VL 72B on HF, Llama 4 Scout on Groq).
• /freebird-ask mixes flagship chat models across families so the answers are genuinely independent rather than five Llama 3.3 restatements (Nemotron on NVIDIA, Grok-3 on GH, DeepSeek-Chat on OpenRouter, Mistral Large, Llama 405B on SambaNova, Qwen 72B on HF).
• /freebird-reason uses only reasoning specialists and bumps max_tokens to 8192 to accommodate chain-of-thought.
• /freebird-quick uses only small fast models (8B class) and caps max_tokens at 1024.

To change the per-task picks, edit src/freebird/tasks.py — one dict per task.

CLI reference

Binary: ~/.local/bin/freebird (Python venv at ~/.freebird/venv/).

Fanout output contract

Every freebird fanout run creates a workspace at /tmp/freebird-YYYYMMDD-HHMMSS/:

• prompt.txt, system.txt — the inputs sent to each provider
• <provider>.md — raw streamed output per provider, with a ===FREEBIRD-META=== footer carrying status + token counts
• <provider>.sh — the script each tmux pane runs
• summary.json — the only file Claude should read after a fanout — contains an outputs array, each entry with {provider, model, file, status, latency_ms, prompt_tokens, completion_tokens, error_detail, content}

At the end of freebird fanout, the CLI prints:

[fanout] summary: /tmp/freebird-YYYYMMDD-HHMMSS/summary.json

Parse that path, read the JSON, then synthesise — never paste raw provider output to the user.

Config

• Location: ~/.freebird/config.toml (chmod 600).
• Env vars override TOML: NVIDIA_API_KEY, GROQ_API_KEY, CEREBRAS_API_KEY, OPENROUTER_API_KEY, GITHUB_MODELS_API_KEY, CLOUDFLARE_API_KEY (plus CLOUDFLARE_ACCOUNT_ID), MISTRAL_API_KEY, TOGETHER_API_KEY, SAMBANOVA_API_KEY, HUGGINGFACE_API_KEY.

Providers included

NVIDIA NIM, Groq, Cerebras Cloud, OpenRouter (:free pool only), GitHub Models, Cloudflare Workers AI, Mistral La Plateforme, Together AI (-Free models), SambaNova Cloud, HuggingFace Inference.

Deliberately excluded (user pays a subscription): OpenAI/Codex, Anthropic/Claude, Google Gemini, Moonshot/Kimi.