v0.3.0
In 0.2.0 one agent at a time did things. In 0.3.0 several agents can work together on one request, agents can work on your own folders, and everything you add to Astros installs one way, as a plugin, and runs sandboxed.
Several agents in one conversation
Assign the work — mention two or more agents anywhere in a message and press Enter. One agent takes the lead and splits the request into tasks, the others take their share, and a plan card under your message shows each task, who owns it, its progress and the running cost. Type while it runs to add a steering note; Stop ends it. When the job is only one task, you simply get a reply.
Ask each instead — ⌥Enter sends the message to every agent it mentions, each answering separately. From the search bar, @a @b … offers Plan with A + B and Ask A + B separately.
Agent profiles — click an agent's avatar or name in a reply to see who it is, how it is set up and what it has done in this conversation, with a link to its settings.
Agents that work in your folders
Working directories — give an agent folders on its Preferences page, each Read or Read & write. A conversation starts with its agent's folders as chips above the message box; add, remove or change them there. Another agent in the conversation asks before it touches one (Allow read, Allow read & write or Deny, for this conversation only).
Edits you can undo — a file an agent changed shows in the reply with its +/− counts. Open shows the diff, Revert puts the file back, and Undo revert brings the change back again. Claude Code and Codex agents start in the first folder on the list.
Plugins
One way to install — + Install… in Preferences installs a plugin from a folder, a URL, a catalog, npm or GitHub. A plugin can bring agents, models, tools, MCP servers and skills; before it installs, Astros shows what it can reach and which accounts it will ask for, and an update asks again only when it wants more. A skill package from npm or GitHub installs the same way. Plugins in Preferences lists everything installed, built-ins included.
Sandboxed — installed plugins run isolated and can reach only the sites they declared. On macOS and Linux, local MCP servers run sandboxed too, limited to the folders and sites you approve in the add dialog; servers launched with npx, uvx and similar are pinned to the version you approved. On Windows a one-time Enable isolation step turns this on, and local MCP servers wait until it is done.
Tools you can call by name — ElevenLabs is now a set of tools (speak, dialogue, sound effects, music, transcribe, change voice, isolate voice) any agent can be given, and @elevenlabs.say Good morning runs one straight from the search bar. Luma, Kling, Stability and Leonardo arrive the same way, and @media uses whichever of them you grant it.
In the conversation window
- Queue a follow-up — press Enter while a reply is still running and your message waits its turn. Edit, Send or remove it from the strip above the message box.
- Drafts stay put — a half-written message stays with its conversation when you switch, and typing anywhere in the window lands in the message box.
- Agents can ask you — an agent can stop mid-reply to ask a multiple-choice question.
- Longer runs — an agent can take up to 90 tool rounds per reply (adjustable in Preferences), and a reply that hits the limit gets a Continue button that picks up where it stopped.
- Files — ⌥⌘B opens a conversation's files; right-click one to reveal, open or copy its path; ⤢ shows a preview full screen. Click an image, or ⤢ on a video, for a full-window viewer with zoom and ←/→.
- Images and PDFs — you can now attach them for GPT and other OpenAI-compatible models too, and tools can hand images and PDFs back to the model.
In the search bar
- Browse folders by keyboard — Tab on a folder in
@filesputs you straight in its tree; Tab or → on a file opens it beside the tree. @filesoperators —^srcmatches a folder or file name starting withsrc, and!testexcludes matches.- Parquet files preview as tables.
Models and web search
@claude.haiku runs Claude Haiku 5.5, @gemini.gemini-4-argon adds Gemini 4 Argon, DeepSeek gains a reasoning-effort setting, and ElevenLabs speech moves to Eleven v4. Web search is now a tool you grant to the agents that should search; Claude and Gemini use their own built-in search, and a reply shows each search where it happened. With no default chat model set, Astros asks you to choose one instead of quietly using Gemini.
Also
- Preferences is reorganized: a Your Astros group lists Agents, LLMs, Tools, Servers, Skills, Harnesses, Mentionables and Plugins, each with its own list.
- Write a skill without leaving Preferences: Skills › + Add skill › Write manually….
- A file you attach that macOS won't let Astros read now says why, and where to allow it.
- Astros's local database moves to PostgreSQL 18; existing data migrates on first launch.
- Plugin authors: plugins now declare their parts in
manifest.jsonand export them withdefineMentionable/defineTool; the olddefineAgentis gone.
Multi-agent conversations and installing plugins are Pro. Sync is now its own subscription, separate from Pro; your first sign-in starts a 30-day trial of both.
Upgrading on Windows: your local MCP servers won't start until you run the one-time Enable isolation step that Astros offers.
macOS (Apple silicon and Intel) and Windows. If you run 0.2.0, the in-app updater will offer this build.