The engagement in one paragraph
A leading investment research firm owns deep proprietary financial data — ratings, earnings estimates, financial statements, filings, and a large analyst research library. Its customers increasingly ask questions inside Claude, ChatGPT, and Copilot instead of opening a terminal or a website. Node8 designed and built a production MCP (Model Context Protocol) server on Google Cloud that puts that data inside those assistants, then ran the distribution phase: submitting the connector to the Anthropic, OpenAI, and Microsoft directories and working through enterprise Copilot governance with the firm’s IT team. The published case study covers the business story: Turning Proprietary Data Into an AI Advantage. This page is the hub for the engineering detail.
The problem: the data was the asset, the delivery model wasn’t
The firm had already tried the obvious move — building its own AI chat product — and learned the hard lesson most data owners learn: standalone chat loses to frontier assistants. Users won’t leave the assistant they already pay for to use yours, no matter how good your data is.
MCP inverts that. Instead of competing with Claude and Copilot, your data makes them better, and you meet users where they already are. The strategic case — why an MCP server beats both a chat product and a traditional API program for this goal — is laid out in Why Expose Proprietary Data via MCP Instead of a Traditional API?.
The approach: structured data first, in weeks not quarters
Node8 ran the engagement in phases, deliberately narrow at the start:
- MVP on structured data. The first release exposed three task-oriented tool groups: article and commentary history by ticker (100+ content categories), company snapshot data (overview, earnings estimates, surprise history), and analyst report retrieval. A second batch — fundamental ratios, sales estimates, price history — followed once the pattern was proven.
- Authentication as a product decision, not an afterthought. Full OAuth 2.1 with a hosted identity layer (WorkOS), Google and Microsoft social login for individuals, and per-organization SSO via Microsoft Entra for enterprise customers. Users connect with one click; the firm keeps rate limiting and governance on its side.
- Distribution as its own workstream. Once the server was live, the work shifted to getting it listed: connector submissions to Anthropic, OpenAI, and Microsoft, plus enterprise rollout mechanics inside the Microsoft 365 admin portal.
- Unstructured content scoped as phase two. The analyst research library gets a retrieval layer on Google Cloud Vertex AI Search, driven by the firm’s rich structured metadata.
Timeline that actually happened: kickoff to production server in roughly four weeks; directory submissions underway by week six.
The architecture, briefly
The server runs on Google Cloud with managed serverless infrastructure and automatic SSL. The auth design splits into two planes: user-facing OAuth (the assistant’s one-click connect flow) and machine-to-machine auth between the MCP server and the firm’s internal APIs, with the user’s identity passed through on every call so the firm can rate-limit per user. Tool design was the other big decision — the firm has hundreds of internal API endpoints, and mapping each to an MCP tool would have produced an unusable surface. Node8 grouped them into a handful of task-oriented tools instead.
Every one of those decisions, including the trade-offs and the sequencing traps (DNS before HTTPS before OAuth registration), is documented in How to Architect a Production MCP Server on Google Cloud.
The rollout: three directories, three different processes
Getting listed in the Claude, ChatGPT, and Copilot connector directories turned out to be three genuinely different processes with different prerequisites — Anthropic requires a Claude Team account to even submit, OpenAI routes through its platform dashboard and needs the right org permissions, and Microsoft’s process is form-based with sparse documentation. The full playbook, including what we’d do differently, is in How to Get Your MCP Connector Listed in Claude, ChatGPT, and Microsoft Copilot.
Enterprise distribution has a second half that most teams miss: even a listed connector has to be enabled, governed, and rolled out inside each customer’s Microsoft 365 tenant. Node8 sat with the firm’s IT team in the actual admin portal and worked through it — including a direct “connect to MCP server” path that let the firm pilot internally before marketplace approval landed. That’s covered in Enterprise Copilot Agents and Connectors: What IT Admins Need to Know.
Results so far
- A production MCP server live on Google Cloud, exposing the firm’s core structured datasets through the open MCP standard.
- One-click OAuth connection working across Claude, ChatGPT, Copilot, and Perplexity.
- Connector submissions in flight across all three major directories, with an internal Copilot pilot running ahead of marketplace approval.
- A repeatable enterprise onboarding pattern: create an organization in the identity layer, add trusted domains, configure SSO against the customer’s identity provider — the same steps for every future enterprise customer.
- Phase two (Vertex AI Search over unstructured research) scoped and ready.
Where to go next
- Strategic case: Why MCP Instead of a Traditional API?
- Architecture: Production MCP Servers on Google Cloud
- Distribution: Getting Listed in Claude, ChatGPT, and Copilot
- Enterprise IT: Copilot Agents and Connectors for Admins
- Buyer questions: MCP Servers for Enterprise Data: FAQ
Work with Node8
Node8 designs and builds production MCP servers end to end — architecture, OAuth, Google Cloud infrastructure, and the connector-directory rollout that turns your data into distribution. If you’re sitting on proprietary data your customers would want inside their AI assistant, talk to us.