Skip to content

v0.8.0

Latest

Choose a tag to compare

@github-actions github-actions released this 17 Nov 12:27

The desktop app is updated too: Desktop v0.0.9. This release is about connecting things: MCP servers with one sign-in, your old chats with their projects, and six more chat apps. Plus an offline mode for when nothing should leave your computer.

Sign in to MCP servers

No more copying keys. Sign in once in your browser and the agent gets the server's tools.

/mcp auth monid      # opens the sign-in page
/mcp signout monid   # signs you out
  • A catalog to start from. Add server now opens a list of ready servers: Exa, Firecrawl, Apify, Bright Data, Tavily, Brave Search, Monid and more.
  • Monid gives the agent paid live data: X and Reddit posts, leads, company data, SEO. Find it in the catalog and sign in.
  • Choose how much each server asks. Every call, only writes, or never.
  • Use your Atomic Agent memory in other tools. atag mcp-server shares your memory and skills with Claude Code, Codex or Cursor, read-only by default.

Offline and private

  • Offline mode. One switch, and nothing goes out except to your local model. Press o in /privacy. It shows what was sent and what was blocked.
  • Incognito chats. /incognito saves nothing, and the chat is gone when you leave.
  • Your keys stay put. Keys can move to your system keychain, and programs the agent starts no longer see them.

Bring your chats over

atomic-agent import now keeps each chat in the project or folder it had before. New sources: the Claude app export and another Atomic Agent. Claude Code and Codex imports work the same way. Run it twice, get no duplicates.

Talk to it from more apps

  • Six new chat apps: Slack, Signal, Matrix, iMessage, Microsoft Teams and Google Chat. Each stays off until you turn it on.
  • Telegram replies now stream into one message, and /cron manages scheduled tasks from the chat.
  • Phone push without a bot: atomic-agent push on.

In the chat

  • /rewind goes back to an earlier message. /btw asks a quick side question. /why shows what the last step did. /resume finishes a reply a restart cut off.
  • @ adds a file, /edit opens your editor, /vim turns on vim keys.
  • The agent can ask you a question as a card with numbered options.
  • Next-step suggestions, off by default: Tab takes one.
  • Export a chat to Markdown, HTML or Word, or share it as one redacted file.

Models and spending

  • A daily spend limit for cloud models. At the limit the agent stops, or moves to your local model.
  • New providers: Azure OpenAI, Amazon Bedrock, Google Vertex AI, W&B Inference, Lemonade, Foundry Local, and a model server on another computer on your network.
  • Local models: pause and resume downloads, see which version fits your computer and how fast it runs, and test models on your own recent tasks.

Also new

  • Edit Word and Excel files, create PowerPoint decks, read EPUB and email files.
  • Speech to text with whisper.cpp, no Python needed.
  • New browser tools, a polite site crawler, and code search by structure.
  • 17 new starter skills, such as audit, pre-push-check, meeting-notes and seo-audit.
  • Scheduled tasks can be paused, and can start when a file appears or an email arrives.
  • Fusion workers can ask the planner a question, and show their steps, files and cost.

Fixed

  • "Allow always" on a web request that sends data could allow it for every site. Now it is saved for one site.
  • Codex refused to edit files, saying it was read-only.
  • With offline mode on, a cloud model failed with a confusing error.
  • Fusion did not count all of its spending.
  • "Allow in this chat" for one MCP tool allowed the whole server.
  • Some dangerous git and shell commands ran without asking.
  • A failed step could be retried after its tools had already run.
  • A wrong API key showed a technical error instead of saying the key was refused.
  • The terminal downloaded a bigger local model without asking.
  • The chat hung when the local model was not running. Now it says how to start it.
  • Ollama failed on long chats, and Groq's size error was unclear.
  • Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2 were missing from the Claude subscription list, and some Claude models failed on temperature.
  • Re-reading a file after a long chat was mistaken for a loop.
  • overnight and /goal stopped or kept going without saying why.
  • Answering the folder trust question typed into the message box.

This release was put together by @sosidudku1, who shipped most of it. Thanks also to everyone who reported these bugs: @agrofubris, @nassimus26 and @QBGROUP on GitHub, and hiago, Bay6, gerius, Sakrael, Majime and The Great Teacher on Discord.

Updating

The app offers the update and asks you to restart. Or run atomic-agent update.