# Control plane

The AgentRun console and API operate on the same agents and runs. Use the console to inspect work and intervene; use the API or CLI to connect it to your application.

## Find the work that needs attention

| Console area | Use it to |
| --- | --- |
| Mission Control | See agent/run counts and recent work. |
| Agents | Create definitions, inspect revisions, choose the current revision, and archive agents. |
| Playground | Run an agent with an input and a selected revision. |
| Runs | Filter by agent or status, follow a timeline, read results, and inspect files. |
| Settings | Configure model keys, tools, and remote MCP servers. |

An `input_required` run may need a person's answer, a budget decision, or a new turn after stopping. Inspect its pending reason before deciding what to do. An error and an unknown cost should remain visible rather than being counted as success or zero spend.

## Control a run

From the run page, follow the timeline, answer questions, or send a follow-up. **Stop** ends the current turn; **Cancel** ends the run. Files and prior messages remain available for inspection.

For automation, the same operations use `/v2/runs/{id}`, `/timeline`, `/messages`, `/stop`, and `/cancel`. [Agents and runs](/docs/agents-and-runs) explains which states can continue.

## Connect tools through MCP

In **Settings → MCP servers**:

1. Add the server's name, address, and authentication method.
2. Choose **Preview** to inspect its tools.
3. Save the server.
4. Edit an agent's tools and grant the specific tools it needs.

A granted tool is named `mcp:<server>/<tool>`. Saving an MCP server does not grant every agent all of its tools. AgentRun keeps the connection credential on the server.

A coding agent can also control AgentRun itself through its `/v2/mcp` endpoint. That connection uses the AgentRun API token; start with `whoami` to discover what this server supports. See the [server skill](/docs/server-skill) for connection instructions.

## Keep changes traceable

Agent definition edits create revisions. Learning changes have their own history and links to the runs that produced them. Runs retain their selected revision and learning snapshot.

Archive finished or retired work to remove it from default lists without deleting its history. Use the archived filter when you need to inspect it again.

Current source also supports [schedules](/docs/server-reference#schedules), [webhook triggers](/docs/server-reference#webhook-triggers), and [run notifications](/docs/server-reference#webhooks). Check the running server's OpenAPI document before building against these routes.

## Console access and API access

A local server asks its owner to set a password on first use. The browser then uses an owner session. Scripts and the CLI authenticate with the server token; they do not need the console password.

The self-hosted server is an owner-operated console, not an organization-wide role system. If you expose it behind a proxy, configure `AGENTRUN_PUBLIC_URL` for the address people open and use HTTPS. Keep its persistent data and credentials outside agents' workspaces.

Continue with [Use the API](/docs/platform-api) or the full [console reference](/docs/server-reference#the-console).
