Tura
Tura is an open-source agent runtime harness that replaces a typical ReAct agent's repeated model round trips with one runtime-managed command graph, aiming to use fewer tokens while matching or beating baseline success rates on coding benchmarks.
Tura restructures an agentic coding task so the model does not have to re-enter after every tool result carrying the system prompt and a growing context, the way a ReAct architecture does; instead, a single planning step compiles the workflow into a command graph that the runtime executes deterministically without further model round trips. On a published comparison of 20 DeepSWE tasks against Codex CLI, run three times per agent, Tura's Direct configuration used 77.5% fewer aggregate tokens at a comparable verifier success rate, while its Balanced configuration reached an 80.0% success rate, 16.7 percentage points higher than Codex CLI, while still using 31.1% fewer tokens.
The published results cover the tested configurations only; broader comparisons across other model providers, local providers and operating systems are described as still on the project's roadmap. It provides both a GUI and a TUI supporting multiple concurrent sessions with rich HTML text.
It is written in Rust and licensed under AGPL-3.0.