Connector engine expansion + MCP server
The execution engine gained deep handlers for the core connectors, and OpFlow shipped an MCP server so an agent can build and run bots conversationally.
Two big things in one window: the connectors moved from catalog metadata to executable operations, and OpFlow became something an AI agent can operate directly.
Deep connector handlers
Server-side handlers were implemented for Google Sheets, Gmail, Google Calendar, Google Drive, Podio, Airtable, RingCentral, OpGrid, and the universal HTTP node — turning the operation catalog into real, runnable steps.
Honesty: these handlers are implemented and executable, but a full end-to-end verification pass against each live API (and per-connector credential wiring) is the next milestone. Until an operation is verified against its real API, treat it as implemented, hardening — not GA.
MCP server (v1)
OpFlow now ships a Model Context Protocol server at /api/public/mcp — streamable HTTP, JSON-RPC 2.0, authenticated with a bearer token (opflow_… keys minted in the admin console and stored hashed).
- 11 tools: list_bots, get_bot, create_bot, update_bot, validate_bot_graph, run_bot, list_runs, get_run, list_node_catalog, get_node_fields, list_connections.
- Claude — in Code, Desktop, or Cowork — can build, validate, run, and inspect bots conversationally. No incumbent offers a native AI creation surface.
AI node
The AI node defaults to Claude (Anthropic API), with plain-text (generate_text) and forced-JSON (generate_json) modes that drop straight into the flow.