Megastructure Arc product history
A permanent record of what changed in every Megastructure Arc desktop release.
Arc 2.2.3 keeps your room layout steady during routine refreshes, rejects malformed replace-all values in file edits, and makes AI settings names and output-token counts clearer.
Arc 2.2.2 doubles the default output room per step for API agents to 16,384 tokens, lets a cut-off reply grow to 32,768 tokens when the provider lists no maximum, and recognises a reply cut off at the output limit even when OpenRouter reports it as an ordinary tool call.
Arc 2.2.0 makes the agents Arc runs cheaper to keep working and easier to get unstuck. Any of them can start from the lean tool list, each turn shows how much of its input the provider's cache served, and an agent can read up to eight files in one call. Long-running agents send far less old text to the model after each checkpoint and keep the file hashes they need to save their work. When a provider never answers a step, the recovery card can now discard it, and Retry unfinished step now really retries. Simplified and Traditional Chinese now call an agent by one name. Fixes include Claude Cowork's restart prompt, servers that say “Streaming is not supported”, provider account ids in room errors, room access for agents limited to some rooms, arc mcp --no-ensure with no hub running, arc stop --storage, and claims, which now survive a hub restart.
Arc 2.1.9 brings desktop stability improvements for long-running agent tasks.
Arc 2.1.8 smooths your first hour with Arc. A new install opens in Start here, and once an agent is in the room, Next steps offers the moves that make Arc pay off. Connecting a coding app and activating now take one click. Connect your agent asks you to restart an app before you send the join phrase, and once your agent joins, tells you to say “Check Arc” in that app after you post. A newly joined agent is asked to tell you which room it joined, errors for a wrong room code or a closed Arc say what to do, and screenshots agents take of Arc go to a folder you and your agents can open. Arc also no longer breaks a project's Rust toolchain while it checks which tools are installed.
Arc 2.1.7 keeps agents in step with you and with each other. Arc can wake an idle Claude Code session when a message, @mention, task or reminder arrives for its agent. Agents Arc runs now re-read their last 12 exchanges instead of 6, a number you can set for each agent, and carry a short summary of older ones; with Read the room on, an agent woken by a room message first reads what the room said since its last turn. From this release on, Arc saves a copy of the database before an update changes its structure, and harness cards in Settings → Harness adapters can warn about harness versions known to cause problems. Harness cards also keep the version Arc detected.
Arc 2.1.6 gets images to more agents, more reliably. An API agent added from a connection now checks once whether its model sees images, and Agent settings gain an Images choice. A model file loads its vision projector, so it can see images and, in a Full Access room, use the computer. Images that MCP service tools return are saved in the room, Reply delivery becomes a choice for Chat Completions API agents, and agents can fetch built-in steps for reviewing a screen. Timeout and dropped-connection errors now point to Retry unfinished step now, and fixes cover limits changes, traffic-limit edits, agent starts that held up the hub, and redirected command-line output on Windows.
Arc 2.1.5 makes API agents steadier on long and troubled requests. They now stream replies by default, so a long reply that keeps producing output is no longer cut off at five minutes, and when a provider is rate-limited, overloaded or briefly failing, Arc retries only the unfinished step without running finished tool work again. A failure that names a setting opens that agent's setting directly, and Agent settings now show where each value comes from.
Arc 2.1.4 runs Meta's Muse Spark and Muse Wonder as agents over the Responses API, and lets Arc's own agents use local MCP servers that you add or import and grant to them. Muse Code connects to Arc like the other coding apps, and a command-line agent started in a linked folder joins that folder's room on its first Arc call. Add agent is now a source browser, with pins, archiving and Tidy up for saved setups, and API agents no longer hang when a provider holds a request open.
Arc 2.1.3 lets you choose behaviours for new agents as you add them, gives an Arc-managed agent a folder of its own, and shows when a model provider is slow to answer. Claude Desktop's Arc entry is now named arc-cowork, so conversations in the Claude desktop app's Code tab can keep their own Arc identities. Long notes are kept instead of refused, work that stops making progress ends itself, and stopping the local hub no longer takes minutes.
Arc 2.1.2 gives every agent a Behaviours list: what wakes it and what it does, from hearing the room to handling a #tag or checking in on a timer. A coordination editor routes matching messages to an agent or a sequence of Arc commands, with compact receipts. Room working folders are separate from automatic joining, and agents finish accepted work before a folder changes.
Arc 2.1.1 brings images into room conversations and gives managed agents web search and page reading on any model. Vision agents see shared images, screenshots and static mock-ups become room evidence, and every text steer reaches the agent in order.
Arc 2.1.0 lets agents follow the room with editable Listening guidance and a shared turn budget. One picker brings together model sources and presets, agents can chat before choosing a folder, and changing models keeps their text conversation. Model queues, current steps and steer delivery stay visible.
Arc 2.0.9 lets self-hosted vision agents use the desktop and gives loops editable prompt lists. Computer actions follow the model’s measured coordinates, replies reach every asker, and conversations stay steady as loop status changes. Usage now shows reported tokens instead of dollar estimates.
Arc 2.0.8 adds multi-machine project sync and Windows computer use. Capsules continue a project rather than making another copy, clones find their rooms, and Mac agents can reach Arc while it is hidden or minimised. Codex uses one connection, and adapter edits preserve your config files.
Arc 2.0.7 is a fixes release. On a Mac, agents can keep driving Arc while other windows cover it. Each Claude Code conversation keeps its own Arc identity, and a project moved to another computer arrives with its open tasks and every agent. Settings stays one click away with the sidebar hidden.
Arc 2.0.6 is a fixes release. Changing an agent’s model keeps its name and its limits, settings you save reach a working agent from its next turn, and replies between agents reach whoever asked, including after Retry. Agents working from a Full Access room can also change a room’s access.
Arc 2.0.5 lets agents that Arc runs wake each other with a message, so they can hand work back and forth, and lets an agent bring in any model you could pick yourself. Agents can do more on Arc’s own screen, a new presentation mode hides private details for screenshots and demos, and the People tab always shows what a room has spent.
Arc 2.0.4 lets agents move a project’s rooms, decisions, skills and memories between hubs, read and use Arc’s own screens and take screenshots of it, and keeps every read small so local models are not flooded. Busy agents keep their tasks, and a stopped agent can be started again from its menu.
Arc 2.0.3 lets a local agent keep its own notes across checkpoints and stops it from skipping a page it read just before one. Agents also keep their Arc identity when an install restarts their connection.
Arc 2.0.2 lets a local agent continue a long file where it left off after reaching its step limit, shows Running on CPU when a local model is not using the GPU, and makes the Windows installer’s reinstall page clearer.
Arc 2.0.1 keeps your current version installed if you cancel a manual upgrade on Windows, and brings back the Start menu shortcut that the update to 1.9.9 could remove.
Arc 2.0.0 lets Arc-managed agents whose models can see use your Mac in Full Access rooms, adds Stop and Delete to the chat, connects Claude through Anthropic’s own API, and recognises self-hosted models that can see images.
Arc 1.9.9 lets Arc-managed agents set reminders and wait for a task, a background job, or a timer, adds a Thinking level per agent, and keeps self-hosted models streaming through long replies.
Arc 1.9.8 adds repeated work cycles for ongoing goals, editable loop instructions, and easier access to each agent’s settings.
Version reporting update
Arc 1.9.6 gives replies the full conversation width and prevents managed agents from posting the same final answer twice in one turn.
Arc 1.9.5 makes empty rooms easier to understand, keeps message recipients tied to the room, and fixes file references, folder linking, and conversation scrolling.
Arc 1.9.4 gives agents more room to work by making coordination and conversation history leaner. It also strengthens rooms, model runtimes, and the desktop, with focused fixes for updating and protected folders on macOS.
Arc 1.9.3 shows when an Arc-managed agent is working and makes the first message in an empty room easier to deliver. Choose an agent, link a folder, and keep your request with you through setup.
Arc 1.9.2 makes it easier to connect an agent, find your room’s files, and pick up where you left off. New rooms offer clear choices for bringing your own agent or using an Arc-managed model.
Arc 1.9.1 makes the work inside a room easier to finish and refine. You can correct your own messages, read and extend large artifacts without loading them all at once, archive old versions, and save files directly into Arc. Previews can take over the full window with familiar zoom controls, and complete open-source notices are available inside the app.
Arc 1.9.0 makes durable memory easier to find, refine, and govern across a whole project. Agents can update, move, copy, and search artifacts without hauling every body into context; rooms gain defaults, real archives, and duplicate protection; and managed seats can retry, rewind, edit, or exclude remembered exchanges with a complete audit trail.
Arc 1.8.9 turns longer agent work into a durable run you can inspect, steer, resume, and finish cleanly. Agents can browse and preview what they build, work with pasted images, and use a clearer connection system for hosted and local models. Shared local-model servers use memory more efficiently, while stronger process custody and recovery keep browsers, previews, files, and managed turns dependable across restarts.
Arc 1.8.8 opens straight into a room, the way a chat app opens on a conversation rather than a front page. Agents can read text-based PDFs sitting in your workspace, a page at a time, and say plainly when a document is a scan they cannot read. Room permissions are now one honest switch, Safe or Full Access, instead of a per-agent matrix. And an agent can steer the desktop itself: it can open the room, artifact, or message it wants you to look at, so showing you the work no longer means telling you where to find it.
Arc 1.8.7 makes the room transcript the whole story. A prompt you send to an agent now appears in the timeline in your own words, directed messages name who they are for, and one message sent to several agents reads as a single row instead of repeating. Sending to an agent that is mid-turn no longer bounces: the message waits in a visible queue and delivers itself the moment that agent is free. You can also stand up a whole team from the composer, or let an agent do it for you.
Arc 1.8.6 refreshes the desktop with the same cool blue and warm gold visual language as the Megastructure website. Dark and light themes now share clearer hierarchy, calmer surfaces, and more consistent controls. The Composer also grows with a multi-line draft, then caps at a practical reading height so longer prompts stay visible without taking over the room.
Arc 1.8.5 keeps model and workspace setup inside the live room. Choose or repair the room project folder, add reusable local models and self-hosted OpenAI-compatible servers, and move an existing agent between API and local sources without losing its identity. Runtime, credential, and desktop-daemon compatibility fixes make those routes dependable after restart.
Arc 1.8.4 gives every AI provider a named Connection, keeps stable agents separate from the sources that run them, and makes room authority explicit. Long local-model turns now recover when a provider counts more context than Arc predicted, while source receipts keep the active model route honest.
Arc 1.8.3 makes coordination more honest under real agent workloads. Skills now learn from observed use and task outcomes, operators and peers can add context to a running turn without cancelling it, and harness failures identify the actual recovery step instead of blaming the model.
Arc 1.8.2 settles on one set of words for agents and AI models, renders every artifact where it lives, and lets a fitness verdict carry evidence a reader can check. Adding agents and downloading artifacts now end exactly where you expect.
Arc 1.8 makes adding and managing agents substantially easier. Create a room team from a configured model in one compact flow, while clearer source and instance roles, safer inheritance, and more accurate memory keep coordination dependable.
Arc 1.7 makes managed-agent work durable across daemon restarts and long-running tasks. Resume interrupted sessions with their transcript, file history, and work context intact, while stronger completion checks and safer background commands help agents finish cleanly.
Arc 1.6.8 adds an explicit Trusted Host mode for local and hosted agents. It can run bounded PowerShell or shell commands with the logged-in user’s authority, while Safe Workspace remains the default.
Arc 1.6.6 makes file drag-and-drop and picker attachments immediate by default, while preserving the safeguards that keep workspace files intact.
Arc 1.6.5 lets teams move selected rooms and coordination history as portable capsules, and makes it easier to connect, test, and switch external, local, or hosted agents from one guided flow.
Arc 1.6.1 warns before likely credentials are sent, adds operator retraction for room and direct messages, and keeps current room state reliable through reconnects and storage changes.
Arc 1.6.0 lets you drop files directly onto a specific agent in the composer and keeps longer Local AI work moving when context fills.
Arc 1.5.0 lets you create reusable agent profiles, seat multiple agents in a room, and choose the right runtime without losing identity, instructions, or room history.
Arc 1.4.3 keeps the full People, Work, and Receipts sidebar one click away in every room.
Arc 1.4.2 improves the startup experience with a focused Home and easier access to recent rooms.
Arc 1.4.1 fixes a secure-connection problem on Mac.
Arc 1.4.0 brings rooms, cloud agents, and Local AI into one calmer conversation surface with one composer and settings that stay out of the way until needed.
Arc 1.3.0 accelerates bundled Local AI on Windows with Vulkan, fixes Mac launch, and makes first-run settings safer.
Arc now ships the llama.cpp runtime with the app — pick a model and a folder and your local AI is running. Files it mentions in chat open as safe previews with provenance.
Arc Desktop can now run a fully local AI agent: your own model, on your machine, working inside a single project folder — with every change reviewable and undoable.
Arc can now connect supported coding agents to the right room automatically, show live coordination in the desktop app, and notify you when work needs attention.
Arc updates now travel through Megastructure’s update service while retaining signature verification inside the app.
A focused reliability release that removes the empty flash between the boot splash and Arc’s first real screen.
Arc gained a signature-verified updater for Windows and macOS, plus release-signing and notarization hardening.
The first commercial Arc desktop release: a local hub for durable agent memory, tasks, decisions, artifacts, skills, and handoffs.