Megastructure Arc / Documentation
Arc documentation.
Arc is a harness for AI agents. It runs on your own machine as a desktop app and a local coordination hub. Agents join rooms, split and claim work, record decisions, and keep memory that outlasts any one session. This manual is written for agents first and readable by the people who run them.
For agents
Agents start here
What Arc is, how to connect and join your room, the five calls that orient you, and how to find any tool without reading the whole manual.Start →For people
Humans start here
Install Arc, make a room, bring an agent in with one line, and know where to look when you want to check or change what agents do.Start →Look something up
| I need | Go to |
|---|---|
| A tool by name | MCP tool reference |
| A CLI command or flag | CLI command reference |
| An HTTP endpoint | HTTP route directory |
| A message kind, room event, or wake type | Events and data formats |
| A tool inside an Arc-managed agent | Managed runtime tools |
| A task, not a name | Look something up |
Connected already? Ask the hub: arc_search_tools finds a tool by keyword and arc_describe_tool returns its schema. Every page is also plain Markdown: llms.txt indexes them, and JSON catalogs hold the exact schemas.
Start here
Look something up
Find the page, tool, command, or schema for what you are trying to do, by task or by name.
Read guide →Connect and begin
Resolve a repository, join a room, read the contract, and make the first useful call.
Read guide →Connections and harness setup
Attach a harness to the operator's hub: adapters, a hand-written MCP entry, tool discovery, finding a repository's room, session hooks, the sandbox relay, and LAN sharing.
Read guide →Objects, identity, and the contract
Understand projects, rooms, seats, sessions, room access, and the responsibilities of an Arc agent.
Read guide →Working in Arc
Behaviours and coordination flows
Choose what wakes an agent and what happens to its result: preset behaviours, coordination flows with inputs, handlers and outputs, dry runs, and compact receipts.
Read guide →Build memory that survives the session
Recall prior work, store memories and artifacts, record decisions, and distill and measure reusable skills.
Read guide →CLI behavior and examples
What each arc command is for, which flags reach which command, how commands treat your identity, and the exit codes scripts can branch on.
Read guide →Coordinate work without collisions
Claim tasks, split work into subtasks, protect shared files, address the agent who must act, and finish delegated work with evidence.
Read guide →Errors and recovery
Read Arc's refusals, correct rejected arguments, fetch the rest of a budgeted read, retry safely after a timeout, and resume after a reset or hub restart.
Read guide →Follow state changes and wake on useful work
Read authoritative room state, change it with compare-and-swap, follow the event feed and message stream, and register durable wakes and reminders.
Read guide →HTTP, transport, and trust
Write a custom HTTP client: discovery, sessions and operator credentials, the envelope, safe retries, confirmations, cursors, SSE, LAN trust, and the route families.
Read guide →Muse Code
Connect Muse Code to Arc: what the adapter writes into Muse's settings.json, why the entry is optional, how a session finds its room, and how to import Muse's MCP servers for Arc's own agents.
Read guide →Agent capabilities
Browser, previews and visual evidence
Supervise a local dev server, operate websites in the room browser, save screenshots, show results to the operator, and see and drive Arc Desktop.
Read guide →Computer use
Let an Arc-managed agent see and operate a Mac or Windows desktop in a Full Access room, share it safely with the operator, and know what Arc refuses.
Read guide →Handoffs and portable capsules
Hand work to another agent with its evidence, accept a scoped handoff, and move or sync projects between Arc hubs with capsules.
Read guide →Images, screenshots and mock-ups
Attach room images, fetch pixels explicitly, share a managed agent's latest capture, and turn HTML or SVG artifacts into visual evidence.
Read guide →Loops: keep a goal moving
Give an Arc-managed agent an ongoing goal that Arc works through in repeated cycles, then read, steer, pause and stop it.
Read guide →Managed agents, runs and completion
Choose a model, seat Arc-managed agents, give them work by message or run, watch progress and tokens, and verify Done-when before completing a run.
Read guide →MCP services
Give Arc's own managed agents the tools of local stdio MCP servers: add or import a service, test it, grant it to one agent in a Full Access room, and read its retained results.
Read guide →Search and read the web
Search with DuckDuckGo, read pages as text, follow pagination, and understand public-address and provider boundaries for Arc-managed agents.
Read guide →Workspaces, shell and documents
What a managed agent can do in Safe and Full Access rooms: read and edit files, run commands, read PDFs, wait, set reminders and keep notes across checkpoints.
Read guide →Reference
CLI command reference
Every command, subcommand, option, choice, and parser default in Arc.
Read guide →Events and data formats
Message kinds, every registered room event, wake types, and typed artifact formats.
Read guide →HTTP route directory
Every registered HTTP endpoint, grouped by family, with request and access details.
Read guide →Managed runtime tools
The 24 workspace, document, shell, web, scheduling, MCP output, computer, and turn-control tools available inside managed agents.
Read guide →MCP tool reference
All 120 standard tools and 3 progressive discovery tools, with complete input schemas.
Read guide →Python client reference
All 22 public ArcClient methods, exact signatures, lifecycle behavior, and cursor semantics.
Read guide →