Desktop v0.0.5

The desktop app gets safer approvals, Telegram pairing in the app, a tidier chat and zoom, plus many fixes. It includes everything in Atomic Agent v0.6.11.

Approvals

• The card shows what you approve. A multi-line command says "(+N more lines)" and opens its full preview from the start. The "Ran ..." line in the chat says the same.
• Allow in this chat approves the same program (for example git) for the rest of the chat. Other programs and other chats still ask, and nothing is saved.
• Shell commands no longer see your API keys, and MCP servers get only the keys in their own config.

Chat

• Finished turns fold into one summary line, like "Used 4 tools · 12 s". Click it to see the steps. Settings › General can always show them.
• Ask again under the last reply of a turn that ran no tools. After a turn that changed files, a greyed icon explains why it is not offered.
• Zoom: Cmd+= and Cmd+- zoom in and out, Cmd+0 is Actual Size. The sidebar moves to Cmd+\ (Ctrl on Windows and Linux). Right-click on text opens an edit menu with spellcheck.
• Tasks and Skills have their own rows in the sidebar. A narrow window gets a drawer with the chat list.
• New task in plain words: describe what the agent should do and when, and it sets up the schedule in a new chat.
• Command palette (Cmd+K) finds things with typos and missing letters, such as "mdl" for Models.
• The model list in the composer shows every downloaded local model and every cloud provider with a working key.

Studio

• Improve prompt rewrites your prompt with the chat model, with Undo.
• Recent lists your last 12 prompts on each page.
• Greyed-out modes say why they are off. A busy cloud model offers Retry or another model. The video gallery has Load more and a full Pinned tab.

Setup and settings

• Telegram pairing in the app: Pair shows a 6-digit one-time code. Send /start <code> to your bot within 5 minutes. Every Telegram line that needs a restart has a Restart agent button, and the app stays open.
• MCP smart add: paste a server's command line, its URL or JSON.
• Claude Code and Codex subscriptions are in Add provider.
• Memory model card shows its own state: Starting, Running, Not running, Off or Not downloaded.
• Too big for this Mac: downloading or starting a model that needs more memory than the Mac has asks first and suggests one that fits.
• Move to Applications: a Mac app run from the disk image or Downloads offers to move itself, so updates keep working.
• llama.cpp custom server suggests port 9931, the default of current llama-server. Saved addresses are not changed.
• Diagnostics opens on health rows for engine, model and provider, each with one fix button.
• The chosen working folder is remembered across restarts, and running turns keep going when it changes.

Fixes

• The agent restarts by itself if it stops unexpectedly, and the window reloads after a page crash.
• A local model that is not running says so and offers Start model. A model that could not load, or moved to the CPU build, says that instead of asking to download the engine again.
• A server that answers 404 or 405 is named as the wrong server. A rejected key shows "Key rejected" with Restart model.
• "Local model ready" appears in setup only after the model started.
• Thinking budget offers only values llama.cpp can hold, and thinking models get room to finish.
• The embedding model keeps running when chat switches to cloud.
• The uncensored model is labelled "Fewer refusals" with a plain explanation.

Downloads

• macOS, Apple Silicon: Atomic-Agent-0.0.5-arm64.dmg
• Windows, x64: Atomic-Agent-Setup-0.0.5-x64.exe
• Windows on ARM (Snapdragon): Atomic-Agent-Setup-0.0.5-arm64.exe. Local models run on the CPU there.
• Linux, x64 and ARM64: .AppImage or .deb