MCP server

Nolixan ships a Model Context Protocol server. Point any MCP client (Claude Desktop, an IDE, your own runtime) at it and your connected providers' actions appear as tools — each call still passing through the policy gate and into the audit log.

Endpoint

POST /api/v1/mcp/
API-Key: sk_...

It's a JSON-RPC 2.0 endpoint supporting initialize, tools/list, and tools/call. Tools are named provider__action — the same convention as everywhere else:

{
  "jsonrpc": "2.0", "id": 3, "method": "tools/call",
  "params": { "name": "github__list-repos", "arguments": { "per_page": 10 } }
}

A tools/call result returns MCP content blocks with an isError flag — a denied policy or a failed action comes back as isError: true rather than throwing, so the model can react.

Runtime config

GET /api/v1/mcp/config

returns the live endpoint URL, auth method, the environment, the connected providers, and the tool count — so a setup screen doesn't have to hardcode anything:

{
  "mcp_endpoint": "https://.../api/v1/mcp/",
  "authentication": { "method": "API-Key header", "header_name": "API-Key" },
  "connected_providers": ["github", "slack"],
  "tools_count": 87
}

You can scope the exposed tools with the provider or category query params, or to a saved tool-set.

MCP vs native tools

If you're calling an LLM API directly, the agent-framework adapters may be simpler than running an MCP client. See native tools vs MCP.