Files
shumai/docs/developer/mcp-server.mdx
T
Yiling-J b3cae4f2b0 feat: move MCP to /mcp and add teammate operations (#500)
* feat(operations): move MCP to /mcp and add search, comment, metadata, version, and upload operations

- Serve the MCP server at /mcp instead of /api/mcp
- Add search_assets, copy_assets, get_folder_instructions, list_comments,
  create_comment, resolve_comment, list_metadata_fields,
  update_asset_metadata, create_version_stack, list_versions, and
  remove_version on the agent, MCP, and CLI
- Add create_file and create_version (agent, MCP) with Creation Context
  metadata, and complete_upload (MCP) for presigned two-step uploads
- Replace the native agent create_file/create_version tools with
  Operation-backed ones
- Route the comment-complete, metadata, copy, version stack, and
  AGENTS.md read REST endpoints through their Operations

* docs(readme): add built-in MCP server and CLI to features list
2026-10-27 14:37:26 +08:00

113 lines
7.1 KiB
Plaintext

---
title: Shumai MCP Server
description: Let external AI agents such as Claude Code, Cursor, or Kiro work with your Shumai projects.
---
Shumai includes a built-in [Model Context Protocol](https://modelcontextprotocol.io) (MCP) server. Any MCP-compatible agent can connect to it and browse projects, look at images, videos, and documents, create folders, and organize assets on your behalf, using the same tools as Shumai's built-in agent.
<Note>
This page is about connecting **external agents to Shumai**. To connect Shumai's own agents to
other tools, see [MCP Servers](/using-shumai/mcp-servers).
</Note>
---
## Connecting a Client
The server uses the Streamable HTTP transport at:
```
https://<your-shumai-host>/mcp
```
It authenticates with an **API token**:
1. In Shumai, open **Team Settings** → **API Tokens**.
2. Enter a name (for example, `claude-code`) and click **Generate**.
3. Copy the token from the dialog. It is only shown once.
Then add the server to your client, sending the token as a Bearer header.
**Claude Code**
```bash
claude mcp add --transport http shumai https://<your-shumai-host>/mcp \
--header "Authorization: Bearer <your-api-token>"
```
**Cursor, Kiro, and other clients with a JSON config**
```json
{
"mcpServers": {
"shumai": {
"url": "https://<your-shumai-host>/mcp",
"headers": { "Authorization": "Bearer <your-api-token>" }
}
}
}
```
<Warning>
An API token acts as you: the agent can do anything your account can do in Shumai. Store it like a
password, use a separate token per client, and revoke tokens you no longer use from **API
Tokens**.
</Warning>
---
## Available Tools
| Tool | What it does |
| ------------------------- | ---------------------------------------------------------------------------------------------------------------- |
| `list_projects` | Lists the projects you can access, with each project's root folder. |
| `list_assets` | Lists files and folders in a folder, page by page. |
| `search_assets` | Searches a folder and its subfolders by name, by meaning (when AI search is enabled), or by metadata. |
| `create_folder` | Creates a folder. |
| `rename_asset` | Renames a file or folder. |
| `move_assets` | Moves files or folders into another folder. |
| `copy_assets` | Copies files or folders into another folder of the same team. |
| `delete_asset` | Moves a file or folder to the project trash. Marked as destructive so clients can ask you to confirm. |
| `read_asset` | Looks at a file: an image, video frames, PDF pages, a 3D model view, or the text of a text file. |
| `get_asset_render` | Gets the result of a `read_asset` call that was still rendering. |
| `download_asset` | Gets download links for up to 50 files: the web preview version by default, or the original with `proxy: false`. |
| `get_folder_instructions` | Reads a folder's instructions for AI agents (its AGENTS.md). |
| `create_file` | Starts uploading a new file into a folder (see [Uploading Files](#uploading-files)). |
| `create_version` | Starts uploading a new version of a file. |
| `complete_upload` | Finishes an upload started with `create_file` or `create_version`. |
| `list_comments` | Lists the review comments on a file, with their replies. |
| `create_comment` | Comments on a file (optionally at a timecode) or replies to a comment. |
| `resolve_comment` | Marks a comment as resolved, or unresolved again. |
| `list_metadata_fields` | Lists a project's metadata fields, such as status, rating, and custom fields, with their choices. |
| `update_asset_metadata` | Sets metadata values on a file. |
| `create_version_stack` | Stacks files in the same folder as versions of one asset. |
| `list_versions` | Lists the versions in a version stack, latest first. |
| `remove_version` | Takes a version out of its stack. The file is kept. |
### Reading Files
`read_asset` attaches the images to the result so the AI model can see them: up to 8 images per call, each up to 2 MB. Any other images are listed as links only. Every result also includes a link to each image, valid for about an hour. Pass `inlineImages: false` to get links only, for example when your client can't show images.
Rendering video frames or PDF pages can take a while. If it takes longer than about 20 seconds, `read_asset` returns `status: "pending"` with a `taskId`, and the model calls `get_asset_render` with that `taskId` to get the result. Only the user who started the read can get its result.
### Uploading Files
Uploading takes three steps, so large files never pass through the AI model:
1. `create_file` (or `create_version`) returns an `uploadUrl` and an `uploadId`.
2. The agent uploads the file's bytes to `uploadUrl` with an HTTP `PUT`, for example `curl -T video.mp4 "<uploadUrl>"`.
3. `complete_upload` with the `uploadId` creates the file and starts processing it (previews, transcoding).
The upload URL and ID are valid for about an hour, and only you can complete your upload. If the project has **Creation Context** metadata fields (for example, the prompt or model used to make a file), pass their values as `metadata` to `create_file` or `create_version`; `list_metadata_fields` shows which fields they are.
The agent must be able to reach the upload URL. With Shumai's local storage, it points to your Shumai server; with S3 storage, to your S3 endpoint.
Every call runs with your permissions: for example, a **Reviewer** can list assets but cannot create or rename them. Actions are recorded in the team audit log and attributed to you, marked as coming from MCP.
---
## Troubleshooting
- **401 Unauthorized**: The token is missing, mistyped, or revoked. Generate a new token and update your client configuration.
- **Permission errors on a tool call**: Your role in that team or project does not allow the action.