Skip to content

Desktop v0.0.8

Choose a tag to compare

@sosidudku1 sosidudku1 released this 09 Nov 21:45

Comes with the terminal agent v0.7.0.

New

  • Run overnight. In the composer's + menu: give a task, a folder, a budget and a time limit (8 hours by default, up to 24). The budget field is hidden for local models. An action that needs your approval waits for you in the app. A small card in the corner shows the time and spend and lets you open or stop the run. With Telegram paired, you can have the report sent there when the run ends. Keep the app open while it runs.
  • Always allow. The approval card has an Always allow button that saves a rule, for example Always allow "git status" in this project, so the same safe command is not asked again. Saved rules are listed in Settings › Privacy, where you can delete them.
  • What ran without asking. Under a finished turn, "Auto-approved 3 actions" opens the list of what ran without asking and why: your rule, or allowed earlier in this chat.
  • Continue from Claude Code or Codex. Command palette › Continue from Claude Code / Codex, or Settings › Import › Continue a task. Pick the session that stopped; it opens as a new chat with its history and a message to carry on, ready for you to press Enter.
  • The best model for this computer. On the small starter model, the app downloads the best model your computer can run in the background ("Getting the best model for this computer: Gemma 4 26B-A4B QAT, 42%") and asks before switching: Switch now or Later. Chats keep working on the small model meanwhile.
  • The agent shows you where to click (macOS). Arrows, boxes and numbered steps appear over any app. A mark goes away once your pointer reaches its spot and moves on, when you press Esc, or when its time runs out.
  • Attach an app window. The button next to + (or File › Attach App Window…) attaches a picture of another app's window to your message.
  • Fusion workers you can steer. Each running worker has Steer and Stop, shows tokens and cost, and is marked when it looks stuck. A note reaches the worker at its next step; stopping one leaves the others running.
  • Cheap, Balanced, Best. After adding OpenRouter, AI/ML API or Gemini, the first three models are marked Cheap, Balanced and Best.
  • Plugins in Settings. Settings › Plugins shows each installed plugin and its parts, lets you switch them on or off, allow or stop plugin code, and remove a plugin. Installing stays in the terminal (atomic-agent plugin install).
  • Live plan. A plan bar above the composer ("Plan 3/7") follows the agent's checklist and folds to a chip.
  • What a chat costs. A chip beside the context ring shows the chat's cost and your plan limit, with reset times in the tooltip.
  • /status in the composer prints what the running turn is doing.
  • Better Markdown. Nested lists, footnotes, safer links; raw HTML shows as text.
  • Tool calls from MCP servers read as words, for example "Used github: create issue".
  • The side panel lists Steps and Reasoning newest first.
  • Long chats stay fast: only the newest part is drawn ("Show N earlier messages"), and pictures off screen leave memory.
  • Fallback notice. When your main provider is rate-limited or down and you set a fallback, the chat says it switched.

Fixes

  • Long sessions: the agent restarts cleanly before the system runs out of memory instead of being killed mid-work, and crash reports say why it stopped.
  • Linux: the app starts on Ubuntu 24.04, both the .deb and the AppImage, without libfuse2.
  • Linux: AppImage installs now get in-app updates. The update feed now carries the Linux files it was missing.
  • Opened from Finder or the Dock, the app finds CLIs such as claude and codex in ~/.local/bin, Homebrew, npm, Bun and Volta folders (macOS, Linux).
  • Windows: model search and the context-window chip no longer fail after the agent printed its list and then crashed on exit.
  • A local server's HTTP 400 now says what to change (the model does not take tools or images, the context is full) instead of "Couldn't get a reply".
  • Entering Fusion no longer picks an embedding-only provider as the chat model.
  • DO_NOT_TRACK=1 is honoured.
  • Studio's Improve prompt works with Kimi K3.1 and GPT-5 instead of failing with HTTP 400.
  • Catalog vision models now also download the file they need to read images, and the list says "text only" when that file is missing.