How does Codebuff actually work?

Codebuff runs multiple agents, each tuned for a specific task.

The main agent ("Buffy") runs on Claude Opus 4.5. It reads your prompt, gathers context, and spawns subagents. The orchestrator is available in several variants:

• base2 - Default mode orchestrator
• base2-lite - Lite mode (faster, cheaper)
• base2-max - Max mode (best-of-N selection)
• base2-plan - Plan mode (no file writes)

• File Picker (Gemini 2.0 Flash) - finds relevant files
• Code Searcher - grep-style pattern matching
• Researcher (Grok 4 Fast) - web and docs lookup
• Thinker (GPT-5.1, Gemini 2.5 Pro) - works through hard problems
• Editor (GPT-5.1, Claude Opus 4.5) - writes and modifies code
• Reviewer (Claude Sonnet 4.5) - catches bugs and style issues
• Commander (Grok 4 Fast or Claude Sonnet 4.5) - runs terminal commands

In Max mode, Codebuff spawns multiple editors with different strategies. A selector compares the outputs and picks the best one.

1. Tree-sitter scans your repo and builds a code map
2. File pickers and searchers find relevant code
3. Thinkers analyze the problem if needed
4. Editors generate changes
5. Reviewers check for issues; commanders run tests

The server is stateless. It streams requests to model providers (Anthropic, OpenAI, Google, xAI) over websockets. Your code stays local; only relevant context is sent.