MCP Basics
This page is the simple mental model for MCP in ConvEngine.
What MCP does
MCP lets the engine decide and call tools during a chat turn.
Basic flow:
- user asks a question
- planner chooses
CALL_TOOLorANSWER - tool output is stored in
context.mcp.observations - planner eventually writes
context.mcp.finalAnswer - response resolver generates final user output
Core tables
MCP table map
| Table | Purpose | Used by |
|---|---|---|
| ce_mcp_tool | Tool registry and scope | McpToolStep / ToolOrchestrationStep |
| ce_mcp_db_tool | DB SQL template fallback config | McpDbToolExecutor |
| ce_mcp_planner | Planner prompts by intent/state | McpPlanner |
Two execution modes
When each mode is used
| Mode | Trigger | Main output |
|---|---|---|
| Planner mode | Natural chat request (no direct tool_request) | context.mcp.observations + context.mcp.finalAnswer |
| Direct mode | inputParams.tool_request provided | inputParams.tool_result |
Start here for testing
Use planner mode first for real chat behavior. Use direct mode only for deterministic tool testing.