Automating AI Integration: MCP Server Pilot for a Leading Global Financial Data Provider

Our Client

The client is a leading provider of transparent and independent credit ratings, benchmarks, analytics, and data that support decision‑making across global capital, commodity, and automotive markets.

Challenges

• Secure data monetization at scale: Expose data as MCP tools for internal/external consumers (AI agents / external LLMs), with scalable API-driven distribution tools.
• Tool onboarding & governance complexity:: large catalog growth (tools derived from metadata + whitelisting). Manual tool coding doesn’t scale; requires metadata-driven generation and controlled.
• Operational readiness (DevSecOps + observability: Integrate CI/CD, monitoring, and security controls  while keeping deployments cloud-native and manageable (Cloud Run initially).

Tech Stack

Framework: FastMCP, Python, Agent/LLM: Google ADK, Gemini 2.5 Pro (Vertex AI), Data: BigQuery, DevOps/Platform: Docker, GitHub Actions, GCP Cloud Run

Solution

• Introduce metadata-driven tool generation: Generate MCP tools from BigQuery metadata tables. Standardize tool definitions using a JSON Intermediate Representation (IR) so metadata extraction and code generation can evolve independently.
• Build a decoupled tool factory pipeline: From Metadata to JSON IR → JSONIR to MCP Tools → Register MCP tools in base MCP server, where generated  MCP tool implementation code that is loaded/served by the MCP server.
• Automate end-to-end updates via GitHub Actions: Generate Python tools, update the base server repo, containerize, and deploy the MCP server to Cloud Run.

Results

• End‑to‑end MCP pilot proven on GCP: Delivered a working MCP server flow where the ADK agent discovers and invokes tools over Streamable HTTP, executing parameterized queries against BigQuery and returning JSON results, with Gemini 2.5 Pro as the baseline model for tool selection.
• Automated tool generation: Server deployment pipeline in place: Implemented the pipeline pattern which produces JSON IR, generates tool code, and the base MCP server loads generated tools; with automation hooks designed for GitHub Actions → container build → Cloud Run deployment.