Tura

About Tura

Tura is an open-source coding agent and harness framework designed for automated software development tasks. It is installable via npm (tura-ai) or by cloning its GitHub repository, with install scripts provided for macOS, Linux, and Windows. The project publishes documentation and a blog alongside its codebase.

The harness operates without tool calling, context-compaction turns, or skills, instead using CLI macros and decision trees to combine patching, compiling, validation, and log inspection into single LLM turns. It applies backward reasoning from outcomes to root causes, manages its own context window and runtime prompt via CLI commands, and follows a test-driven approach that investigates and reproduces an issue before making a patch.

Tura is positioned against Codex CLI using a published benchmark of 348 long-horizon sessions covering DeepSWE, rewrite, and design tasks (20 DeepSWE tasks, 5 rewrite tasks, 2 design tasks, averaging 43.5 agent turns per session, 15,138 total agent turns). Reported results show up to 83.1% fewer turns on the rewrite benchmark and up to a 16.7 percentage point improvement in DeepSWE pass rate compared with Codex CLI; the rewrite cohort is evaluated using 236 stable multi-item checks rather than a single binary score. The benchmark is described as community-contributed, allowing others to submit their own tasks and evaluation reports. It is intended for developers working on long-running development workflows rather than isolated model capability or simple Q&A comparisons.

Key features

• Open-source coding agent harness with no tool calling, context-compaction turns, or skills
• Applies backward reasoning from outcomes to root causes
• Manages its own context window and runtime prompt via CLI commands to reduce irrelevant context
• Investigates and reproduces the full execution path before making a patch
• Community benchmark where anyone can contribute benchmark tasks and evaluation reports

Frequently asked questions