MCP Runtime Flow

How Agents Should Use the Four MCP Tools Together

A well-designed agent workflow using Solid MCP typically follows this pattern:

User Question
     │
     ▼
(Optional) Semantic Model QA — discover which model covers the question
     │
     ▼
(Optional) Glossary — understand business terms in the question
     │
     ▼
SQL Generation — get the grounded SQL query
     │
     ▼
Agent executes SQL on DWH using end-user credentials
     │
     ▼
(Optional) Asset Information — get more detail about specific assets in the results
     │
     ▼
Answer returned to user

See Getting Started with the Solid MCP Server for the parameters and examples of each tool.

End-to-End Runtime Flow

This is the full flow when an AI agent calls Solid's MCP server to answer a business question. Each step is labeled with its security implications.

Step 1: Business question
         Business user asks a question via their agent/chat interface.
         (e.g., "How many green loans were promoted last year?")
         │
         ▼
Step 2: Request SQL from Solid
         Agent formats and sends an MCP tool call to the Solid MCP server.
         The agent authenticates to Solid via the user's identity/access layer.
         Solid verifies the user has MCP access before serving any model.
         │
         ▼
Step 3: Test SQL using DB execution state
         Solid's SQL generation engine identifies the relevant semantic model(s),
         generates candidate SQL, and sends it to the Solid DB Agent for validation.
         (Security note: DB Agent runs as read-only service principal;
         scoped to only exposed schemas; applies LIMIT + timeout)
         │
         ▼
Step 4: Run on DB; Pass/Fail (+error) passed back
         DB Agent executes the validation query on the data warehouse.
         Returns only Pass/Fail status + error details if failed.
         No query results flow back through Solid.
         Correction loop runs if failed (up to 3 attempts).
         │
         ▼
Step 5: Run response returned
         After a Pass, the validated SQL is ready to return to the agent.
         │
         ▼
Step 6: SQL for agent
         Solid returns the validated SQL to the agent, along with:
         - An explanation of the reasoning (which model, which terms resolved, how interpreted)
         - Optional semantic context (glossary terms, field meanings)
         │
         ▼
Step 7: Agent runs SQL on DWH using end-user permissions
         The agent executes the SQL directly against the data warehouse
         using the end-user's own credentials and permission set.
         (Security note: RLS, column-level security, table permissions — all enforced
         by the warehouse as if the user had queried it directly.)
         │
         ▼
Step 8: Business answer
         Query results are returned to the agent, which composes the answer
         and presents it to the business user.
         (Data never traverses Solid — only the SQL does.)

The critical security property: at no point does actual query result data flow through Solid. Solid generates and validates SQL; execution and data retrieval happen directly between the agent and the data warehouse under end-user credentials. See Security Architecture for more on this boundary, and When Does Solid Execute SQL? for how this compares to the different execution boundary used during ingestion and benchmarking.