OpenViking Control Plane — MCP Server + CLI

MCP server and CLI for the OpenViking control plane (topapi) — manage OV libraries (Collection). Both front-ends share one core (client.py), so a tool added once is available from MCP and the CLI alike.

Covers the 6 core control-plane Actions:

Endpoint

The control-plane TopAPI is compiled into the OpenViking data-plane cluster; each Action is served by the data-plane gateway at:

{endpoint}/api/openviking/{Action}
# default endpoint: https://api.vikingdb.cn-beijing.volces.com/openviking
# full URL e.g.: https://api.vikingdb.cn-beijing.volces.com/openviking/api/openviking/ListOpenVikingCollections

The Action lives in the path (no ?Action=&Version= query). The request body is the Action's params (e.g. {"ResourceID": "..."}).

The default endpoint points at the reserved public data-plane gateway (not open to traffic yet). For local testing set --endpoint / VIKING_ENDPOINT to a kubectl port-forward, e.g. http://localhost:18080.

Authentication

The only method: an Ark AgentPlan ApiKey, sent as an Authorization: Bearer <key> header on every request (the backend's authorizeControlPlaneByArk reads the key only from this header — it does not accept X-API-Key). Auth is pluggable (common/auth.py → BearerTokenAuth); an AK/SK signer can be swapped in later without touching the rest.

⚠️ Write actions like create require the account to have AgentPlan deduction activated, otherwise they return ProductUnordered. Read-only actions (list/get/usage/delete) are not gated.

Configuration

CLI usage

uv sync                      # or: pip install -e .

# set the key once via env, then drop the per-command flag
export AGENTPLAN_API_KEY=ark-xxxxxxxx

# read-only
uv run ov-cp list
uv run ov-cp get   <ResourceID>
uv run ov-cp usage <ResourceID>
uv run ov-cp api-key <ResourceID>

# create (consumes paid quota; with source=agentplan only --name is needed —
#         model names default, and the model ApiKey falls back to the configured key)
uv run ov-cp create --name my_kb

# delete (irreversible)
uv run ov-cp delete <ResourceID> --yes

Flags override env. The endpoint defaults to the public gateway; override it only for testing (e.g. against a port-forward) with -e / VIKING_ENDPOINT — uv run ov-cp -e http://localhost:18080 list. ov-cp --help works without any config.

MCP usage (stdio / uvx)

The server defaults to stdio transport, so it can be launched as a subprocess by any MCP client. Add to .mcp.json:

{
  "mcpServers": {
    "openviking-controlplane": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/volcengine/mcp-server#subdirectory=server/mcp_server_openviking_controlplane",
        "mcp-server-openviking-controlplane"
      ],
      "env": {
        "AGENTPLAN_API_KEY": "ark-xxxxxxxx"
      }
    }
  }
}

For local development point it at your checkout instead:

{
  "mcpServers": {
    "openviking-controlplane": {
      "command": "uv",
      "args": ["run", "--directory", "/abs/path/server/mcp_server_openviking_controlplane",
               "mcp-server-openviking-controlplane"],
      "env": {
        "AGENTPLAN_API_KEY": "ark-xxxxxxxx"
      }
    }
  }
}

Run with SSE instead via mcp-server-openviking-controlplane --transport sse.

Agent skill

A Claude Code / agent skill that documents the ov-cp workflow lives at skills/openviking-controlplane/SKILL.md. Symlink or copy it into your agent's skills directory (e.g. ~/.claude/skills/) to let an agent drive the control plane.

⚠️ create_collection / delete_collection create/destroy billable resources and are exposed as MCP tools; their descriptions instruct the model to confirm with you first. Rely on your client's tool-permission prompt as the final gate.