Skip to content

MCP Servers

AlphaBase can connect to MCP servers for external tools and context providers.

Manage MCP Servers

Use CLI:

Terminal window
alphabase mcp

Or manage through runtime routes/UI controls depending on your client surface.

Config Schema

MCP servers are defined under the mcp config key, in two forms:

{
"mcp": {
"local-example": {
"type": "local",
"command": ["/absolute/path/to/reviewed-mcp-server"],
"environment": {
"API_KEY": "{env:SOME_API_KEY}",
},
},
"remote-example": {
"type": "remote",
"url": "https://mcp.example.com/mcp",
"headers": {
"Authorization": "Bearer {env:REMOTE_MCP_TOKEN}",
},
},
},
}
  • type: "local" — spawns a subprocess; command is an array containing the executable followed by its arguments, with optional environment, enabled, and timeout.
  • type: "remote" — tries Streamable HTTP and then SSE; supports url, optional headers, optional oauth, enabled, and timeout.

See Config.

Typical Flow

  1. register MCP server config
  2. enable/disable per environment
  3. grant permissions for MCP tools
  4. verify server health and tool discovery

Security

After an AlphaBase update, an older OAuth connection may ask you to authorize again. Reconnect it through Integrations and confirm that the authorization page belongs to the service you trust. Other saved projects and settings remain.

For a pre-registered OAuth client with a secret, set oauth.issuer to its trusted authorization server URL. Ask the service administrator for that URL; do not substitute the MCP server URL unless the service documents them as identical.

  • treat MCP servers as untrusted code and data-access boundaries
  • scope credentials minimally and never place secrets in prompts, skill files, or committed configuration
  • review tool permissions (ask, allow, deny) for each MCP namespace
  • require explicit user approval before an MCP tool can expose data or perform a consequential action
  • pin trusted server versions where possible and review updates before enabling

MCP configuration is not a substitute for sandboxing. A connected server may execute code, access external systems, or return untrusted tool descriptions.

See Permissions and Tools.