An experimental, high-performance AI coding agent & harness optimized for GPT models.
Important
Coming Soon: Capybara Code is currently under initial preparation and internal development. The source code, installation packages, and pre-built binaries have not been uploaded or published yet. Stay tuned for the upcoming initial open-source release!
Capybara Code is an open-source AI coding agent designed to build the most optimal harness and toolset specifically engineered for GPT models.
While base model capabilities are essential, Capybara Code investigates how much performance, reliability, and precision can be maximized through harness engineering—combining a rich Terminal UI, an isolated Rust execution sidecar, transactional file mutations, and multi-agent workflows.
- 💻 Interactive Terminal UI (TUI): Real-time task logs, live token/context tracking, cost estimation, and status indicators.
- ⚡ Rust-Powered Sidecar: Safe, transactional file editing with optimistic concurrency, diff reviews, and rollback checkpoints.
- 🔑 Dual Provider Support: Direct OpenAI API integration (
capy auth api) or official ChatGPT/Codex plan support (capy auth login). - 🤖 Sub-Agent Workflows: Task delegation across specialized roles (
explore,executor,architect,reviewer,test). - 🔌 Extensible Ecosystem: Support for Model Context Protocol (MCP) servers and reusable
SKILL.mdworkflows. - 🔍 Context Engine & Repository Map: Smart symbol search, instruction parsing (
AGENTS.md), and automated context selection. - 🤖 Headless & Automation Mode: Run tasks non-interactively (
capy run --jsonl) with machine-readable JSONL event streams.
Note
Source code and build scripts will be made publicly available upon the initial source release.
Capybara Code was created to explore a single premise: How far can we push AI agent performance through harness and tool optimization rather than relying solely on base model updates?
Key Objectives:
- ⚡ GPT-Optimized Harness & Tools: Build the most seamless, high-precision toolset and runtime environment tailored specifically to extract maximum performance from GPT models.
- 🚀 Outperforming Industry Leaders: Surpass the agentic coding performance, accuracy, and task-completion rates of state-of-the-art tools like Claude Code and OpenCode.
- 🧠 Advanced Agent Workflows: Push the boundaries of automated coding through robust planning, multi-agent orchestration, atomic verification, and intelligent context management.
We warmly welcome all forms of community involvement!
- 💡 Feature Ideas & Suggestions: Have a thought on improving our harness design or tools? Share it with us!
- 🐛 Bug Reports & Issues: Found a problem or edge case? Let us know so we can fix it.
- 🔀 Pull Requests & Feedback: Contributions of any size—from docs to code—are greatly appreciated.
★ Star or watch this repository to receive updates when the initial source code and developer preview are officially released!
Licensed under the Apache License 2.0.