a native desktop app for agentic coding. it does not wrap an API or reimplement a harness. it spawns the official CLI and gives it a real interface: agents running in parallel on isolated branches, the model's reasoning streamed live, and a usage guard that acts before you hit a limit instead of apologizing after. built in rust and tauri, so it stays out of your way.
split panes and themes are table stakes. the field ships them too. these are the parts that took real work.
most tools show you a meter. yume acts on it: it declines to start a turn at 99% of your 5-hour window, and reroutes an expensive turn to a cheaper model at 98% of the weekly bucket. it reads your usage through the CLI itself, never your credentials, and fails open if it cannot.
rather than inheriting a blanket bypass, yume replaces it with per-session tool control across eight categories, MCP isolation that ignores repo-local server configs, and bash deny rules evaluated before execution. the one thing it cannot do, stop a specific tool call once a turn is already running, is stated plainly in the docs rather than buried.
the underlying CLI ships constantly. yume keeps a per-version review log reaching back to CLI 2.1.113, which is the unglamorous work that decides whether the integration still functions next week.
every feature is available in the demo. the paid tiers lift the workspace limits, and that is the only difference.
because shipping and supporting software is the fastest way to stay honest about what agentic systems can actually do. yume is built with the same agent loop it exposes, so every rough edge in agentic coding reaches our issue tracker before it reaches a client.