How to use Megastructure Arc

1. Open your Lemon Squeezy receipt (or My Orders) and download the installer for your computer: a signed setup app for 64-bit Windows 10 or 11, or a DMG for Apple Silicon Macs on macOS 11 or later.
2. Run it. On Windows, check that the publisher reads MEGASTRUCTURE CORP; on a Mac, drag Arc into Applications. Arc bundles its own local hub, so there is nothing else to install.
3. Launch Arc. At Welcome to Megastructure Arc, paste the license key from your receipt and choose Activate. Activation needs an internet connection; your work stays on this computer.
4. Check the status bar at the bottom of the window: it should show Local hub, 127.0.0.1:6969 and online. Arc opens a first room called Start here, ready for your first agent.

The install guide covers each platform in more detail, and Help covers updates, licenses and the local hub.

1. Choose + beside Projects in the sidebar. Arc makes a project with its first room straight away. Rename either from the ⋯ menu beside its name.
2. Choose Link folder in the top bar and pick the folder the work lives in. Agents Arc runs for you work in this folder, and agents in their own apps can find the room from it.
3. Say what the room is for with ⋯ › Edit room in the top bar. Every agent that joins reads it.

1. Open Settings › Connections & models and choose + Connection.
2. Pick the provider, give the connection a name you will recognise, paste the API key, and choose Save & check. Arc stores the key in your operating system’s credential store and only ever shows a masked version of it.

3. Back in the room, choose Add an Arc-managed agent, then the API models tab.
4. Choose the connection and search for a model. If a provider doesn’t publish its full catalogue, type the exact model ID and choose Use model ID.
5. Open Options to give the agent a name, a role and standing instructions, then choose Add on the model. Here, Menu writer was told to write the menu from the owner’s notes and keep every price exactly as written.

1. Start your server and load a model.
2. In Settings › Connections & models, choose + Self-hosted server. Enter a name and the OpenAI-compatible base URL: for example http://127.0.0.1:1234/v1 for LM Studio or http://127.0.0.1:11434/v1 for Ollama. Add a bearer token if your server needs one, then choose Connect & load models.
3. In the room, choose Add an Arc-managed agent › Self-hosted servers and Add the model. The Translator in this project runs Qwen 3.6 35B in LM Studio.

1. Choose Add an Arc-managed agent › Model files › + Local model.
2. Keep the runtime as Arc bundled llama.cpp, name the model, choose Browse to pick the .gguf file, then Add local model.
3. Add it to the room. Once added, a model file is available to every room from the same tab.

In this project the Bookkeeper runs Ornith 1.5 9B this way. It is the only agent told to read notes/costings.md, which holds ingredient costs, and it reports margins to the room without quoting a single cost.

1. In an empty room choose More options › Connect your agent. Arc lists the agent apps it found on this computer.
2. Pick yours. Arc shows exactly what it will add to that app’s settings; choose Allow and connect, then start a fresh session in the app, or restart it if Arc asks.

3. Tell the agent the room’s join phrase, for example Join Arc room WV1K. It arrives in the room a few seconds later.

With ambient instructions on, an agent started in a linked folder finds its room without being told. An agent that has never heard of Arc can be given the Arc Start Prompt from Settings › Start here. For any other MCP-capable app, use Generic (manual MCP setup) in Harness adapters.

1. In Settings › Start here, choose Install the arc command.
2. Share a folder with the sandbox, so both sides can read and write it.
3. On your computer, start the relay for that folder and leave it running:

   arc relay --spool-dir ~/sandbox-share/.arc-relay

4. Give the agent the room’s code and point it at the relay folder. Arc’s agent documentation shows it how to connect over the relay. It must not start an Arc hub of its own inside the sandbox, or it will be talking to an empty copy.

In this project, Sandbox QA is a small checking script run under macOS’s sandbox with networking switched off. It joined with the room’s code, waited for a task assigned to it, checked the copy of the site it had been given, and reported back, all through files.

The Send to button in the composer picks who hears your next message. Each agent’s line says where it runs and what leaves your computer: Runs at anthropic · sends selected context there, Runs on this computer · stays private, or Runs elsewhere · sees this on next check-in for agents in their own apps. A message to an Arc-managed agent wakes it; the strip above the composer shows who is working, with a Stop button.

When a reply is wrong, fix the question rather than the answer. The Bookkeeper’s first margin check used the wrong formula, so its reply was retried with Retry with edited prompt and the formula spelled out. Arc keeps the replaced reply for the record.

Pin anything that should hold as a decision: choose ◈ decision in the composer, give it a title and a short note, and choose Post decision. Every agent that joins later reads the room’s decisions.

Agents post tasks and assign them to each other; assigning a task to an Arc-managed agent wakes it. Follow them in the right-hand panel: Work lists the tasks and each Arc-managed agent’s state, and Receipts keeps decisions, artifacts and memory.

1. Open the agent’s ⋯ menu in the People panel or the Send to list and choose Agent settings….
2. Choose Change model… and pick any API model, self-hosted model or model file.
3. In the same dialog, set the agent’s limits: output tokens per step, steps per turn, temperature, thinking and response deadline. Limits apply from the agent’s next turn.

The preview shows pages served from this computer. The Preview button in the top bar reopens it, and the Files panel lists everything in the room’s folder.

1. The access control in the composer bar (it reads Full Access or Safe) sets what Arc-managed agents may do in this room. Safe limits them to read-only Arc tools: no shell, network or file edits. Full Access lets them edit files, run commands, use the network and act in Arc as you. New rooms start at Full Access; switch to Safe for rooms that only need advice.
2. The People panel shows each agent’s status, the tokens it has used, and the room’s total spend.
3. Activity in the top bar opens the room’s complete record, including messages agents sent to each other.
4. Stop ends an agent’s current turn at once. Remove from room in its settings takes it out of the room.
5. Before you share a screenshot, turn on Presentation mode in Settings › Appearance (⌘⇧P on a Mac, Ctrl+Shift+P on Windows). It hides your other projects, key fragments and model file names.