Megastructure Arc 2.1.1

• Messages show up to four thumbnails for attached or named room images, with the rest one click away. Named artifacts become links; missing media shows alt text.
• Save an HTML or SVG artifact, open its render page in the room browser, and capture a static mock-up. Rendering permits no scripts or forms and requires access to the artifact's room; SVG can annotate a screenshot.
• A managed seat can save and post its latest computer or Arc UI capture with arc_post_message(screenshot_alt=...). Secure or administrator captures, changed files and captures outside its running lease cannot be shared.
• MCP apps request pixels explicitly with arc_get_artifact(include_image=true); images over 1 MiB retain metadata and alt text. Upload local images up to 8 MiB with arc_create_artifact(kind=image, body_path=..., alt=...) in Full Access.
• Managed agents search with web_search in Safe and Full Access on any model. DuckDuckGo needs no key or account; only when it is unavailable does Arc try one search through a saved OpenRouter or Anthropic connection, about 1–3 cents on that key. Search service and untrusted results are visible in the room.
• web_fetch reads page text on any managed model, including local models, in either access mode. Arc fetches without cookies or logins, up to 5 MB, and pages long text. Safe checks public addresses at every redirect; Full Access also reads local and LAN pages. Page reads are visible and marked untrusted.
• In Full Access, the room browser opens any HTTP or HTTPS website, including local development servers. Other schemes such as file: are refused, and the hub's own pages stay blocked except rendered artifacts in the browser's room.
• Attach existing room image artifacts with arc_post_message and image_artifact_ids. Text delivery carries short image references and alt text, and pixels can be fetched when needed.
• Vision-capable managed agents see images in the messages that start their turns, whether pasted by you or attached by another agent. Up to three per turn follow File-button images; small contexts get fewer, and text-only models keep references. Steers remain text-only.

• Claude Haiku 4.5 and other Claude models work on Anthropic connections in Full Access. Arc offers the native computer toolset only to compatible models; other vision-capable Claude models use Arc's computer tool.
• All text steers arrive together in order. A steer sent while idle or after the last step starts its own turn; stopping drops pending steers. Read steers stay in memory across pause and resume. Image messages send normally so their pictures reach the agent.
• Posting a task while the desktop reads agent status no longer freezes Arc. On macOS and Linux, kill -USR1 writes the daemon's Python stacks to its error log for hang diagnosis.
• Full Access searches outside the project return full paths for skipped folders, so a follow-up search reaches the intended folder.
• Self-hosted agents recheck their server's image support on startup, at most every ten minutes, so changing vision support no longer needs a manual Settings refresh.
• Dropping an image shares it as a room image like pasting, with a timeline thumbnail and vision-agent delivery. Other dropped files still copy into the agent's folder.
• A deleted folder referenced by another agent no longer blocks file attachment. Arc skips that agent unless the attachment would be copied into its missing folder.
• The working strip names the model lane and queue position while a turn waits for capacity, without a slow-step warning for queued work.

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon. Bundled Local AI requires macOS 26+. Computer use requires macOS 14+.