llm-cli-gateway API reference

llm-cli-gateway exposes the Model Context Protocol over local stdio by default and over Streamable HTTP at /mcp when the operator starts llm-cli-gateway --transport=http.

Machine-readable API files

• OpenAPI 3.1 description
• RFC 9727-style API catalog
• MCP server card
• Runtime-derived MCP tool index

Authentication

HTTP transport supports a configured static bearer token and optional built-in OAuth 2.0 with PKCE. A shared static bearer token is one shared principal, not multi-tenant isolation. OAuth advertises mcp, workspace:admin, and cli:admin. Remote workspace creation or registration also requires LLM_GATEWAY_WORKSPACE_ADMIN=1; remote provider-admin operations also require LLM_GATEWAY_CLI_ADMIN=1.

JSON errors

HTTP transport errors always include error and may add structured fields such as code, retryable, or session-capacity details. MCP protocol errors and tool failures use normal JSON-RPC or MCP tool-result envelopes.

Endpoints

• POST /mcp - initialize and send MCP JSON-RPC requests.
• GET /mcp - continue a Streamable HTTP MCP session with Mcp-Session-Id.
• DELETE /mcp - close a Streamable HTTP MCP session with Mcp-Session-Id.
• GET /healthz - unauthenticated liveness and readiness summary with bounded process-health metrics.