Technology RadarTechnology Radar
Trial

Codebuff is an open-source, terminal-based multi-agent coding CLI that coordinates specialized sub-agents (file picker, planner, editor, reviewer) in a continuous review loop. It claims 61% vs 53% on coding evals vs Claude Code across 175+ real-world tasks, and a 2.5× speed improvement from parallelism.

Why It's in Trial

Codebuff's multi-agent review loop is architecturally distinct from most CLI agents that use a single model in a sequential tool-call loop:

  • Four-agent pipeline: File Picker (identifies which files need changes) → Planner (sequences edits to handle dependencies) → Editor (executes precise changes) → Reviewer (validates output and reprompts the editor if needed). The loop continues until the reviewer is satisfied
  • Self-improving: Stores knowledge of your coding preferences in local markdown files and refines them across sessions — builds a project-specific model of patterns over time
  • Model flexibility: Supports any model available on OpenRouter — not locked to Anthropic like Claude Code
  • TypeScript SDK: Generator-based SDK for composing custom multi-agent workflows, spawning subagents, and integrating into CI/CD pipelines
  • Codebase-aware: Deep repository understanding — structure, dependencies, hidden patterns — enabling style-consistent edits across large codebases
  • Freebuff: Ad-supported free variant with built-in web research and browser use — no subscription or API key required

Benchmark Claims

Codebuff reports 61% vs Claude Code's 53% across 175+ coding tasks on real-world open-source repositories. These figures are self-reported and not yet independently audited — treat as indicative rather than definitive. The 2.5× speed claim is based on parallel multi-agent execution vs. sequential single-agent.

Key Characteristics

Property Value
Interface CLI, SDK
License Open source
Underlying model Any OpenRouter model (Claude, GPT, Qwen, DeepSeek, etc.)
Pricing Free (Freebuff, ad-supported) / Paid (BYOK via OpenRouter)
Sandbox Host terminal (no built-in isolation)
Provider Codebuff
Website codebuff.com
GitHub CodebuffAI/codebuff