What is Bonfire

Bonfire is an AI agent orchestration framework that ships four load-bearing claims most frameworks leave to you:

1. It learns — A local-first LanceDB-backed knowledge graph under knowledge/ accumulates patterns across runs, paired with priority-based prompt compilation that drops lowest-priority context first when the window is tight. Fork the repo and the graph is empty; use it for a month and it knows your code style.
2. It competes with itself — Dual or triple workflows where two or three agents solve the same problem from different angles. A Sage synthesizes the best of both. Not task-splitting — structured disagreement.
3. It feeds its own fire — A pipeline engine where workflows are frozen Pydantic models, DAG-validated, gate-enforced, checkpoint-resumable. Knights write failing tests; Warriors write minimal code to pass them. The pipeline enforces the split.
4. It gets better the more you use it — Per-stage, per-agent, per-token cost receipts (opt-in toggle, default off). Fuel burns are the default surface; raw numbers are there when you want them. The 6-tier XP progression (Spark through WhiteHeat) compounds heat as you use the framework.

Install

pip install bonfire

Numbers

Free tier

The free tier ships the full forge. All eight agents, all seven pipeline stages, full priority truncation, full gamification loop. Free is not a demo — it is the product. Monetization sits above via plugins and tier-gated services, never via engine feature removal.

What Bonfire is NOT

• Not a freemium demo with gated features
• Not a "coming Q3" product — if a claim can't cite source, it doesn't ship
• Not a replacement for your entire stack — it orchestrates agents, it doesn't write your business logic
• Not "zero Python required" — Bonfire is a Python package

See the Limitations page for a complete accounting of what doesn't work yet.