Tura

Tura is an open-source agent runtime harness for coding agents that delivers better results while using fewer tokens than conventional tool-calling agents. It is built for developers running long-horizon coding tasks who want lower token cost and higher task success without switching language model providers. Tura is distributed as a CLI, TUI and GUI.

Instead of making the model re-enter the context after every tool result, Tura turns a task into one runtime-managed command graph, so deterministic execution continues without another model round trip; its macro command_run tool lets an agent build a multi-step execution tree and run related shell commands, patches and tests in a single LLM turn rather than five separate turns.

Tura guides the model to reason backward from an estimated near-final state toward the current one rather than only reasoning forward, which the project reports helped it pass more task verifiers than Codex CLI on a published DeepSWE comparison. It also treats context as part of its runtime state machine, using task_status, runtime prompts and recursive execution manuals so sessions can be renamed, refreshed and managed instead of letting skill files and stale history accumulate.

Tura installs globally via npm install -g tura-ai on macOS, Linux and Windows, or can be built from source with platform-specific install scripts. It does not bundle provider credentials, so a user must configure an LLM provider and choose a model before sending a first prompt. Tura is released under the AGPL-3.0 license, and the published benchmark comparisons do not yet cover every provider, with Anthropic/Claude, Google/Gemini and other configurations listed as part of the documented roadmap.

Use cases

• Fixing a frontend bug by reconstructing the failure state before writing code
• Running a long-horizon coding task such as the DeepSWE benchmark set