2.3 MCP Best Practice
Evaluation & Onboarding Guide
Getting research value from Orbit MCP
Orbit MCP gives any modern AI assistant (Claude, ChatGPT, Copilot, Gemini) a direct conversational line into Orbit's knowledge base: 70M+ financial documents, 55,000+ public companies, and 10 years of history. Every answer is grounded in real filings, transcripts, and news, with citations back to the source documents. Used like a quick search, it returns quick answers. Used like a research analyst you can brief and delegate to, it returns grounded, analyst-grade work in minutes. This guide is how you reach the second outcome.
1. Where MCP fits in the stack
Orbit is a stack, not a single product, and MCP is one surface into it. Choosing the right surface for the job is the single biggest determinant of whether Orbit becomes a nice-to-have or a must-have in your workflow.
Orbit Insight
Full SaaS research UI
Structured, repeatable research across many companies, with saved workspaces and team sharing.
Orbit MCP
Conversational, in your assistant
Ad hoc, opportunistic deep dives on a narrow set of companies or one theme, without leaving your AI workflow.
Agent Marketplace
Pre-built and custom agents
A repeatable task you want to run systematically across many entities or on a schedule.
Knowledge Base API
Raw AI-ready data and feeds
Building your own models or pipelines and you need the clean data layer delivered to your infrastructure.
The architecture in one line
MCP is the conversational front door to Orbit's RAG infrastructure. It handles the opportunistic framing and the deep dive. When you need scale, the same data opens up through Agents (systematic and scheduled) and the Knowledge Base API (your own pipelines). The teams that get the most from Orbit move fluidly between these surfaces rather than forcing one surface to do every job.
2. When MCP is the right tool
Reach for MCP when all three are true:
Narrow question. Roughly 1 to 5 companies, one theme, or one time window.
You want to stay in your assistant, where you draft memos and explore hypotheses.
You need grounded answers with citations, not the assistant's general training data.
Use another surface when you want to:
Screen 1,000+ names on complex logic → Orbit Insight with Agents.
Monitor a portfolio continuously → News Signal Pulse or Portfolio News Tracker.
Extract bulk data for a model → consume the KB via API.
Run the same job weekly → build or clone an agent.
Rule of thumb. If you ask the AI the same MCP question more than three times in a week, graduate it into an agent. It will be faster, more consistent, and shareable with your team.
3. How to ask, with examples
Output quality is almost entirely a function of prompt quality. A strong MCP question names five things: the entity (be specific), the scope of documents, the time window, the analytical lens (change over time, contradictions, peer divergence, tone shift), and the output format you want. The three plays below are the highest-ROI on-ramps we see.
Play 01 — Pre-earnings thesis check (Fundamental analyst)
I own [TICKER] on a thesis of [one sentence thesis]. Read their most recent 10-Q, the last two earnings call transcripts, and any investor presentations from the past six months. Surface every data point that supports this thesis, and separately every one that challenges it. Flag language changes from the prior quarter. Cite all sources.
Play 02 — Temporal delta (The edge over traditional tools)
Read [TICKER]'s risk factor sections in their last three 10-K filings. What risk factors were added, removed, or materially rewritten between FY22, FY23, and FY24? Show before and after language for each change, and flag which ones relate to AI or capex.
Play 03 — Disagreement mining (Where the signal hides)
Compare what [TICKER] management said about demand in their most recent earnings call versus what is disclosed in their latest 10-Q risk factors. Where is there tension between the two? Highlight the specific language on both sides.
Why citations matter for compliance. Every MCP response cites the source documents behind each claim. For MiFID II or SEC regulated teams using AI-assisted research, that source trail is your audit defense. If an answer comes back without sources, it came from the assistant's training data rather than Orbit, so always require them.
4. Invoke agents straight from chat
Agents from the Orbit Marketplace can be invoked directly inside your AI assistant. This bridges ad hoc conversation and the systematic, repeatable logic that lives in agents. The most effective pattern: use a direct prompt to explore and frame, invoke an agent for structured, consistent output, then return to a direct prompt to interpret and decide.
Run the Anti-Greenwashing Checker on BP's most recent sustainability report and prospectus. Give me the structured output, and flag any commitment that is softened or absent across the two documents.
5. Limits, and where to graduate
MCP runs inside your assistant's chat window, so it inherits a finite context budget and is stateless between sessions. Those are not flaws, they are signposts. Each limit maps to a deeper Orbit capability.
Run the same analysis on 500+ companies
Orbit Insight + Agents
Monitor a portfolio continuously
Portfolio News Tracker + alerts
Produce identical structured output every time
Custom agent with concepts
Pipe clean text or vectors into a proprietary model
Orbit KB / API feeds
Share a research workspace with a team
Orbit Insight workspaces
Start in two minutes
Activate your Orbit MCP key via your account manager or the Orbit Insight settings panel.
Add the Orbit MCP server to your assistant (Claude, ChatGPT, Copilot, or Gemini).
Verify: "Using Orbit, summarise the most recent 10-Q for Microsoft with citations." Confirm the answer cites source documents, then run Play 01 on a name you know well.
Orbit MCP Evaluation Guide — Orbit Financial Technology — V1.0 / Confidential
Last updated