Tura is an open-source, local coding agent designed to address the inefficiencies and unreliability of existing AI coding assistants. It targets developers seeking verifiable performance improvements and reduced operational costs, offering a substantial token-budget advantage by minimizing repeated context and model round trips. Tura aims to provide a more robust and cost-effective solution compared to standard LLM-based coding tools.

How It Works

Tura employs a novel command_run macro tool that consolidates multiple small tool calls into a single LLM turn, enabling the agent to build and execute multi-step command sequences efficiently. This approach significantly reduces conversational overhead and token usage. Additionally, Tura utilizes a backward reasoning strategy, guiding the LLM to estimate the preceding state ($s_{n-1}$) before reasoning towards the goal ($s_n$), which aims to improve the quality and correctness of generated code by identifying root causes more effectively than standard forward reasoning. Context is managed as a runtime state machine, using task status, runtime prompts, and recursive execution manuals to keep active context scoped to the current task, preventing degradation from stale information.

Quick Start & Requirements

Installation is available via NPM (npm install tura-ai) or GitHub Packages (@tura-ai/tura). Source checkout and build are also supported via Git clone and platform-specific scripts (install.ps1 for Windows, install.sh for macOS/Linux). Tura requires configuration of an LLM provider and model selection upon first launch; it does not bundle credentials. Documentation is available via a GitBook-style index at docs/SUMMARY.md.

Highlighted Details

• Achieves up to 83.1% fewer turns and 83.5% fewer aggregate tokens compared to Codex CLI High on rewrite benchmarks.
• Improves DeepSWE pass rates by up to 16.7 percentage points, reaching an 80.0% success rate with 49.6% fewer tokens in a balanced configuration.
• Consolidates sequences of inspect, patch, build, and test commands into a single command_run LLM turn.
• Employs backward reasoning to estimate prior states, aiming for more considered and correct code generation.

Maintenance & Community

Contributions are encouraged to be small, reviewable, and supported by test evidence. Detailed contribution guidelines, roadmap priorities (0.1.x stabilization, 0.2 task-planning), known issues, and a Code of Conduct are available in the docs/ directory. Support can be sought via bug reports and feature requests.

Licensing & Compatibility

Tura is licensed under AGPL-3.0-or-later. This copyleft license may have implications for integration into closed-source commercial products.

Limitations & Caveats

The published benchmark results do not establish equivalent quality or performance for every configured provider. Broader measurements across different LLM providers (Anthropic/Claude, Google/Gemini), OpenAI-compatible services, local providers, UI latency, runtime parsing, and cross-OS compatibility are identified as known evidence gaps and part of the documented roadmap.