Tura — local, open-source coding agent

[Yohjisakamoto]

Tura — local, open-source coding agent

Tura is a local, open-source coding agent for developers who are tired of vague skill claims, token-saving extensions with no evidence, and agents that change repositories without understanding them first.

• Source: https://github.com/Tura-AI/tura
• Website: https://turaai.net/
• Published benchmark, methodology, and artifacts: https://turaai.net/benchmark

In published DeepSWE v1.1 comparisons (20 tasks, 60 sessions, GPT-5.6 SOL High), Tura Direct used 83.5% fewer aggregate tokens than official Codex CLI High, with 65.0% vs 60.0% verifier success. Tura Balanced reached 80.0% success while using 49.6% fewer tokens. See the benchmark page for methodology, raw artifacts, and scope.

Disclosure: I’m affiliated with Tura-AI, and AI assistance was used to prepare this post; it was reviewed before publishing.

[haabe]

Thanks for posting this here, and for leading with the benchmark and the raw artifacts instead of adjectives.

The frustration you opened with, vague skill claims and agents that rewrite a repo before they understand it, is pretty much why I started building Mycelium, so it's good to see someone coming at the same thing from the runtime side.

Mycelium sits a step upstream of what Tura does. It's about working out whether something is worth building before an agent writes the code, not how efficiently the code gets written. Different layer, same allergy to claims with nothing behind them.

Two things I'd genuinely find useful, if you're up for it:

• What made you post it in this repo in particular? I'm trying to learn who these discussions actually reach.
• On "understand the repo before changing it": how does Tura decide it understands enough to act? That's a question I keep hitting from the discovery side too.

You don't have to say anything nice about Mycelium. How the two ends of the pipeline each handle evidence is the more interesting thing.