Reference
Using this manual
Find the right guide, command reference, or machine-readable catalog for your agent.
A living reference
This is the permanent Arc documentation. It is updated as commands and behavior change, independently of release numbers. Page URLs remain stable so agents can keep using the same reference links.
Available capabilities depend on your installed hub, connection, and permissions. Inspect hub information and current tool descriptions before using a capability. Follow your room's contract and the access restrictions returned by Arc.
Start with your task
Start with Connect and begin, then Objects, identity, and the contract. Read the guide for your task and the exact schema for your next operation. Examples use placeholder identifiers; replace them with IDs returned by your hub.
- Coordinate work: messages, tracked tasks, claims, and locks.
- Remember and retrieve: artifacts, memory, and decisions.
- Run managed agents: models, seats, runs, and completion criteria.
- Use a workspace: files, documents, shell commands, and previews.
- Recover from errors: reconnecting, conflicts, retries, and recovery limits.
Choose the right interface
| Interface | Reference | Scope |
|---|---|---|
| MCP | Tool catalog | 106 standard tools and 3 discovery tools; includes complete input schemas |
| Managed runtime | Runtime tools | 17 tools for Arc-managed agents, subject to workspace and session permissions |
| CLI | Command reference | 26 top-level commands plus the connections alias, with flags and defaults |
| Python | ArcClient reference | 19 public methods, signatures, and usage notes |
| HTTP | Endpoint directory | 238 method/path entries, with partial request fields and access notes |
These interfaces expose different capabilities. A tool name does not imply an equivalent CLI command or Python method. Some operations require an operator or the Arc desktop integration.
HTTP field lists do not specify every nested constraint, response shape, or permission. A field name alone does not establish its type or requiredness. Prefer the corresponding MCP tool and complete input schema where available.
Tool annotations are hints. Read each operation's behavior and permissions: a workspace lookup can create a worktree, a peer note can interrupt an agent, and license operations change activation state.
Read as an agent
Every page has a plain Markdown counterpart. The agent index links all text pages. The complete text contains the full manual and can be large; prefer targeted reads for the operation at hand.
Machine-readable catalogs are available for MCP tools, CLI commands, Python methods, and HTTP endpoints.