Codebuff is an open-source AI coding assistant that edits codebases through natural language instructions by coordinating multiple specialized agents. Unlike single-model tools, Codebuff splits the coding task across dedicated agents — a File Picker, Planner, Editor, and Reviewer — that collaborate to understand project context and produce precise changes. According to the project's published evaluations, this multi-agent approach achieves a 61% success rate across 175+ real-world coding tasks, compared to 53% for Claude Code on the same benchmark set. ## Multi-Agent Architecture The core design principle behind Codebuff is that different stages of a code modification task benefit from different reasoning strategies. When a user submits a natural language instruction, the system activates a pipeline of specialized agents: **File Picker Agent**: Analyzes the request and identifies which files in the repository are relevant to the change. This agent understands project structure, import graphs, and file naming conventions to select the minimal set of files needed. **Planner Agent**: Takes the selected files and the user's instruction, then produces a structured modification plan. The plan specifies what changes need to be made, in what order, and what dependencies exist between changes. **Editor Agent**: Executes the plan by generating the actual code modifications. Because it receives a structured plan rather than a raw instruction, its output is more targeted and less prone to hallucinated changes in unrelated code. **Reviewer Agent**: Validates the generated changes against the original instruction, checking for logical consistency, style adherence, and potential regressions. ## Model Flexibility via OpenRouter Unlike tools locked to a single provider, Codebuff supports any model available through OpenRouter — including Claude, GPT, Qwen, DeepSeek, and specialized coding models. Users can assign different models to different agents: a fast model for file picking, a reasoning model for planning, and a coding-optimized model for editing. This flexibility allows teams to optimize for cost, speed, or accuracy depending on the task. ## TypeScript SDK for Production Integration Beyond the CLI tool, Codebuff ships a TypeScript SDK (`@codebuff/sdk`) that enables embedding its multi-agent capabilities directly into applications. Developers can create custom agents with programmatic control, integrate Codebuff into CI/CD pipelines for automated code reviews, or build internal tools that leverage the agent coordination layer. The SDK exposes typed interfaces for agent definitions, execution contexts, and result handling, making it straightforward to extend the system with custom agents for domain-specific tasks. ## CLI Installation and Usage Installation is a single npm command: `npm install -g codebuff`. Running `codebuff` in any project directory starts an interactive session where the user describes changes in natural language. The tool indexes the project structure on first run and maintains a local cache for faster subsequent operations. The CLI also supports non-interactive mode for scripting: `codebuff run "add error handling to all API routes"` executes the full agent pipeline and applies changes directly. ## Published Agent Marketplace Codebuff includes a marketplace of community-published agents that extend the default set. Teams can publish specialized agents — such as a migration agent for database schema changes or a security audit agent — and share them across projects. ## Community and Traction With 3,400+ GitHub stars, 428 forks, and 793 dependent projects using the SDK as of March 2026, Codebuff has established a growing ecosystem. The project maintains an active Discord community and accepts contributions under the Apache 2.0 license. The 6,000+ commit history reflects continuous development and frequent releases.