Dirac
An open-source AI coding agent built to cut API costs 50-80% versus other agents by curating context instead of dumping entire files into every prompt.
🔗 Visit Dirac
Description
AI coding assistants can rack up surprisingly large API bills, mostly because many of them re-send large chunks of your codebase to the AI model with every single request. Dirac is an open-source coding agent designed specifically to avoid that waste — it's built to give the AI just enough context to do the job well, which its creators claim cuts costs by 50-80% compared to less careful agents, without sacrificing the quality of the code changes.
Dirac is an Apache 2.0-licensed coding agent for multi-file refactoring and complex codebase work, installable via npm or the VS Code Marketplace. It uses hash-anchored edits (referencing stable content hashes instead of fragile line numbers), AST-native manipulation for language-aware precision, and batches multi-file changes into a single LLM round-trip to reduce redundant calls. It supports autonomous tool use — file operations, terminal commands, headless browser control — and project-specific customization via an AGENTS.md configuration file. It requires models with native tool-calling support and has 1,400+ GitHub stars.
💬 Our review
The short version: if your AI coding agent bills are creeping up because of a chatty agent that keeps re-reading files it already saw, Dirac's whole design is built around not doing that, and the hash-anchored edit approach is a genuinely clever fix for the 'line numbers shifted, edit failed' problem that plagues naive agents.
Against general-purpose coding agents that just pipe your whole repo through an LLM on every turn, Dirac's context curation and AST-native editing are real architectural advantages, not just marketing — the 50-80% cost reduction claim is plausible given how much redundant context typical agents send. The tradeoff is that it requires models with native tool-calling support, so it won't work with every model you might want to plug in, and as a newer project (1,400 stars) it hasn't had the years of real-world edge-case hardening that more established agents (GitHub Copilot Workspace, Cursor's agent mode) have. For cost-conscious teams running agents at scale across many repos, that cost difference compounds fast and is worth the migration; for occasional individual use, the savings are real but smaller in absolute terms.
💰 Pricing
📊 Global score
🤖 AI-enriched data
100% open source (Apache 2.0), installable via npm ou VS Code Marketplace ; coût réel = uniquement les appels API du modèle LLM utilisé
Pros
Réduction de coût API revendiquée de 50-80%
Édition ancrée sur hash (robuste aux décalages de lignes)
Manipulation AST native par langage
Batching multi-fichiers en un seul aller-retour LLM
Cons
Nécessite un modèle avec tool-calling natif
Projet plus jeune, moins éprouvé que Copilot/Cursor
Pas de coût "tout compris" — dépend toujours du modèle LLM choisi