Tura

Tura is a free and open-source (AGPL-3.0) local coding agent built by Yohji Sakamoto and the Tura-AI project. Where most AI coding agents are hosted SaaS with opaque internals, Tura ships as a single tool with three interchangeable interfaces — a CLI for scripting, a TUI for terminal-heavy workflows, and a GUI for developers who prefer a windowed experience — all running locally against the repositories you already have on disk. It can inspect and edit source, run commands against the project, and verify its own results before returning a patch, which brings it closer to a real autonomous coding loop than a chat wrapper. Its most unusual feature is the public benchmark area at turaai.net/benchmark, which openly exposes per-run model rounds, token totals, patches, and verifier outcomes — the kind of raw telemetry most commercial agents keep private. That transparency lines up with the project's technical blog, which analyzes topics like how output-token filtering can quietly force extra agent turns and inflate cost. For developers evaluating AI coding agents on evidence rather than marketing, Tura is one of the few options where you can read the source, replay the benchmark, and self-host end to end without a subscription.