Viking Knowledge Base MCP Server

简体中文

Overview

Viking Knowledge Base MCP Server lets MCP clients such as Claude Desktop, Cursor, Cline, and Trae interact with VolcEngine Viking Knowledge Base. It supports managing and inspecting collections and documents, and searching for relevant knowledge chunks.

Features

• Add a document to a knowledge base collection by URL
• Get document information and processing status
• List documents with cursor pagination
• Get collection information and build status
• List collections in the configured project
• Search a collection with optional document filtering
• Authenticate with either a Viking API key or VolcEngine AK/SK credentials
• Run over stdio or stateless Streamable HTTP

Prerequisites

• Python 3.10 or later
• A Viking Knowledge Base API key, or VolcEngine AK/SK credentials

Installation

Install from PyPI:

pip install mcp-server-knowledgebase

Or install it as a persistent command-line tool with uv:

uv tool install mcp-server-knowledgebase

For local development, clone the repository and install this package in editable mode:

git clone https://github.com/volcengine/mcp-server.git
cd mcp-server/server/mcp_server_knowledgebase
uv pip install -e .

Configuration

Authentication

Configure at least one authentication method:

• API key: set VIKING_API_KEY. Requests use Authorization: Bearer <VIKING_API_KEY>.
• AK/SK: set both VOLCENGINE_ACCESS_KEY and VOLCENGINE_SECRET_KEY. Requests use VolcEngine SignerV4 authentication.

When both methods are configured, VIKING_API_KEY takes precedence. If no API key is set, the access key and secret key must be configured together.

Environment variables

Environment variable	Description	Default
VIKING_API_KEY	Viking Knowledge Base API key; takes precedence when configured	-
VOLCENGINE_ACCESS_KEY	VolcEngine access key	-
VOLCENGINE_SECRET_KEY	VolcEngine secret key	-
KNOWLEDGE_BASE_PROJECT	Viking Knowledge Base project	default
KNOWLEDGE_BASE_REGION	VolcEngine region used for AK/SK signing	cn-north-1
KNOWLEDGE_BASE_TIMEOUT	Upstream request timeout in seconds	30
MCP_SERVER_HOST	Streamable HTTP bind host	127.0.0.1
MCP_SERVER_PORT	Streamable HTTP port; falls back to PORT	8000
STREAMABLE_HTTP_PATH	Streamable HTTP endpoint path	/mcp

Running the server

Run with the default stdio transport:

mcp-server-knowledgebase

Run the published package without installing it persistently:

uvx --from mcp-server-knowledgebase mcp-server-knowledgebase

Or run the module directly from a source checkout:

python -m mcp_server_knowledgebase.server --transport stdio

Run with stateless Streamable HTTP:

mcp-server-knowledgebase --transport streamable-http

The default endpoint is http://127.0.0.1:8000/mcp. Set MCP_SERVER_HOST=0.0.0.0 only when deploying behind a trusted gateway.

MCP protocol and deployment security

The server uses MCP Python SDK 2.x and supports protocol revision 2026-07-28, including stateless server/discover negotiation. The SDK also handles older handshake-based clients. Legacy HTTP+SSE is not exposed because it is deprecated by the 2026-07-28 specification.

Streamable HTTP does not turn the configured Viking API key or VolcEngine AK/SK into client authentication. Protect remote deployments with an authentication gateway or MCP-compatible OAuth, and never expose service credentials to callers.

Available tools

• add_doc: Add a document by URL
• get_doc: Get document information and processing status
• list_docs: List documents with cursor pagination
• get_collection: Get collection information
• list_collections: List collections in the configured project
• search_knowledge: Search knowledge in a collection

add_doc

Add a supported document to a collection by URL.

add_doc(
    collection_name="product_docs",
    add_type="url",
    doc_id="product_guide_2026",
    doc_name="Product Guide",
    doc_type="pdf",
    url="https://example.com/product-guide.pdf",
)

Parameters:

• collection_name (required): Target collection name.
• add_type (required): Currently only "url" is supported.
• doc_id (required): Unique ID containing only letters, numbers, and underscores. It must start with a letter and contain 1–128 characters.
• doc_name (required): Document name containing 1–256 characters.
• doc_type (required): One of xlsx, csv, jsonl, txt, doc, docx, pdf, markdown, faq.xlsx, or pptx.
• url (required): Accessible URL of the document.

Example result:

{
  "collection_name": "product_docs",
  "doc_id": "product_guide_2026"
}

get_doc

Get a document's metadata and processing status.

get_doc(
    collection_name="product_docs",
    doc_id="product_guide_2026",
)

Parameters:

• collection_name (required): Collection containing the document.
• doc_id (required): Document ID.

Example result:

{
  "collection_name": "product_docs",
  "doc_id": "product_guide_2026",
  "doc_name": "Product Guide",
  "doc_type": "pdf",
  "url": "https://example.com/product-guide.pdf",
  "add_type": "url",
  "create_time": 1788220800,
  "update_time": 1788220860,
  "point_num": 53,
  "status": {
    "process_status": 0,
    "failed_code": null
  }
}

process_status values: 0 completed, 1 failed, 2 or 3 queued, 5 deleting, and 6 processing. Fields not returned by Viking are null; additional upstream document fields are preserved.

list_docs

List documents in a collection using cursor pagination.

list_docs(
    collection_name="product_docs",
    limit=2,
    next_token=None,
)

Parameters:

• collection_name (required): Collection whose documents will be listed.
• limit (optional): Number of documents to return, from 1 to 100. Defaults to 100.
• next_token (optional): Opaque cursor returned by the previous call. Omit it for the first page. An empty cursor in the result means all documents have been returned.

Example result:

{
  "collection_name": "product_docs",
  "total_num": 3,
  "count": 2,
  "doc_list": [
    {
      "collection_name": "product_docs",
      "doc_id": "product_guide_2026",
      "doc_name": "Product Guide",
      "doc_type": "pdf",
      "url": "https://example.com/product-guide.pdf",
      "add_type": "url",
      "create_time": 1788220800,
      "update_time": 1788220860,
      "point_num": 53,
      "status": {
        "process_status": 0
      },
      "brief_summary": "An introduction to the product.",
      "total_tokens": 345
    }
  ],
  "has_more": true,
  "next_token": "opaque-cursor-for-next-page"
}

total_num is null when Viking does not provide it. Document entries preserve additional upstream fields such as summaries and token counts.

get_collection

Get information and build status for a collection.

get_collection(collection_name="product_docs")

Parameters:

• collection_name (required): Collection name.

Example result:

{
  "collection_name": "product_docs",
  "description": "Product manuals and release notes",
  "status": 1
}

Collection status values: -1 pending build, 0 building, 1 completed, 2 failed, and 3 changing.

list_collections

List all collections in the globally configured project.

list_collections()

This tool has no parameters.

Example result:

{
  "collection_list": [
    {
      "collection_name": "product_docs",
      "description": "Product manuals and release notes"
    },
    {
      "collection_name": "support_faq",
      "description": "Frequently asked support questions"
    }
  ]
}

search_knowledge

Search for relevant chunks in a collection. An optional document filter can include or exclude matching document field values.

search_knowledge(
    query="How do I reset my password?",
    collection_name="support_faq",
    limit=3,
    doc_filter={
        "op": "must",
        "field": "doc_id",
        "conds": ["account_guide"],
    },
)

Parameters:

• query (required): Search query.
• collection_name (required): Collection to search.
• limit (optional): Maximum number of chunks to return, from 1 to 100. Defaults to 3.
• doc_filter (optional): Object with the following fields:
  • op: "must" to include matches or "must_not" to exclude them.
  • field: Document field to filter, such as "doc_id".
  • conds: Non-empty list of values to match.

Example result:

{
  "result_list": [
    {
      "id": "chunk_001",
      "content": "Open Account Settings and select Reset Password.",
      "doc_id": "account_guide",
      "doc_name": "Account Guide"
    }
  ]
}

doc_id and doc_name are null when Viking does not provide document metadata. A non-null doc_id can be passed directly to get_doc.

MCP client configuration

Example stdio configuration using uvx and a Viking API key:

{
  "mcpServers": {
    "knowledgebase": {
      "command": "uvx",
      "args": [
        "--from",
        "mcp-server-knowledgebase>=0.2.1",
        "mcp-server-knowledgebase"
      ],
      "env": {
        "VIKING_API_KEY": "your-viking-api-key",
        "KNOWLEDGE_BASE_PROJECT": "your-project-name",
        "KNOWLEDGE_BASE_REGION": "cn-north-1"
      }
    }
  }
}

You can instead configure both VOLCENGINE_ACCESS_KEY and VOLCENGINE_SECRET_KEY. If all three credentials are present, VIKING_API_KEY takes precedence.

Troubleshooting

1. Authentication errors
   • Verify the API key or AK/SK credentials.
   • When using AK/SK, configure both values together.
   • Check that the credentials can access the configured project and collection.
2. Connection timeouts
   • Check network connectivity to the VolcEngine API.
   • Adjust KNOWLEDGE_BASE_TIMEOUT when necessary.
3. Empty results
   • Verify the project and collection names.
   • For search, try a broader query or remove the document filter.
   • For document listing, continue only with the exact next_token returned by the previous call.

License

This project is licensed under the MIT License.