# External MCP capability map

Which public Rumata MCP capability to use for each need when working without a development machine, and how it differs from the documentation MCP.

Source: https://docs.rumata.dev/guides/agent-external-mcp
Reviewed: 2026-10-31

## Before you start

- An MCP client connected to Rumata’s product MCP, not only to this documentation site.

<a id="two-servers"></a>

## Two different MCP servers

| Server | What it does |
| --- | --- |
| Documentation MCP at `/mcp` on this site | Read-only. Lists, reads and searches public documentation. It runs no product action and reads no account. See [Documentation for AI agents](https://docs.rumata.dev/guides/agents). |
| Rumata product MCP | Acts on your factory through public tools: create backends, workflows, files, deployments and more, subject to your permissions. |

Use this site to learn the contracts, and the product MCP to perform them. External MCP clients can read these guides without a development machine.

<a id="getting-started"></a>

## Resolve IDs and call safely

1. Call `list_factories` and `get_project_context` to resolve the factory, parent project, existing `databaseConnectionId`/`defaultSchemaName` and `projectStorageId`. Reuse these bindings. Context discovery does not provision resources or return credentials; a null ID means no single visible binding was found.
2. Before an unfamiliar website action, call `get_operation` and inspect its schemas and scope selectors. A top-level `factoryId` does not replace endpoint filters.
3. Give each `call_operation` a UUID `requestId` and reuse it only for identical retries.
4. Search the tool list by domain, path or shared capability name.

> **Names are discovery terms**
> Internal `rumata.*` capability names are discovery terms, not callable aliases. Call only the public tools the server returns, and do not invent callable names or payloads. Report a missing capability instead of silently substituting a weaker implementation.

<a id="delegation"></a>

## Direct work or agent delegation

Respect the user’s choice between doing the work directly and delegating it. For direct work, use the machine operations and website APIs below. Factory initial prompts and coding messages start separate agent work; do not use them just to create an empty workspace.

## Related guides

- [Documentation for AI agents](https://docs.rumata.dev/guides/agents)
- [Agent reference: build on Rumata](https://docs.rumata.dev/guides/agent-reference)
- [Workflows and connectors](https://docs.rumata.dev/guides/agent-workflows)
- [Platform capabilities](https://docs.rumata.dev/guides/agent-capabilities)