MCPforhedgefunds:aresearchworkflowyoucandefendtoyourcommittee
MCP lets your AI assistant call research tools, but it won’t decide what to research, when to run or whether the answer holds up. Here’s how to build a scheduled workflow where every figure links to its source and stale inputs get flagged before anyone reads them.
Key takeaways
MCP connects your AI assistant to research tools. Deciding what to research, when to run and whether the answer holds up is still your job.
Store the source, publication time, reporting period and any calculation next to every number the agent writes, so a reviewer can check it.
Run the workflow by hand before scheduling it, and test three cases: useful results, no results and a source that can’t be reached.
Start with the hosted connection, and build your own MCP server only for a specific gap, such as your own company universe or screen.
Example research workflows on Bigdata.com cost $0.57 to $1.82 a run, and new accounts start with $10 in free credits.
Every small fund that tries AI research hits the same wall. The agent writes a confident paragraph with a revenue figure in it, and someone asks where the figure came from. If the honest answer is “the model said so”, the paragraph is worthless. One fund manager we spoke to put it plainly: “I can’t really present that to my credit committee.”
This post is about getting past that wall. It walks through one workflow, a scheduled update on an investment theme, built so that every number traces back to a document and every stale input is flagged before anyone reads it. The examples use Bigdata.com’s MCP server and two open cookbooks you can copy, but the habits carry over to whatever data you connect.
Pick the job before the tool
Start with something your team already does by hand. For most small funds that’s one of three things: catching up on overnight news across the book, keeping a thematic basket current, or checking whether a story changes a thesis. Bigdata.com has worked examples of portfolio briefings and sentiment scans, plus a custom MCP cookbook for screening companies against a theme. Borrow whichever is closest to your routine.
Then decide who reads the output. A PM catching up at 7am wants a few lines and a link per name. A researcher exploring a theme wants the evidence laid out so they can argue with it. Same data, very different notes, and the agent can’t guess which one you meant. Write down the companies, the sources, the time window and what the finished thing should look like. If you haven’t settled on a connector yet, our guide to financial MCP connectors compares the options.
Write the brief as if a new analyst were picking it up
The instructions do most of the work, and they’re the part people rush. Write them for a capable junior on their first day: which companies, what counts as evidence, what the answer must contain and, the part everyone forgets, what to do when something is missing.
Scheduling is a separate piece. In the dashboard example below, an external automation starts the agent on a timer, and MCP is simply how the agent reaches the data. Keeping those two apart in your head makes debugging much easier later.
A few decisions to make before the first scheduled run:
The cutoff time and timezone. “Daily” means nothing until you say daily as of when.
The search window. A fixed period, such as the last 24 hours, is simplest. “Everything since the last successful run” survives a missed day. A little overlap catches stories that arrive late, as long as you tell the agent what to do with duplicates.
How old is too old. Yesterday’s price and last quarter’s filing are both normal; a price from last week isn’t. Set a freshness rule for each type of input.
Then run it by hand a few times. Before trusting a schedule you want to have seen three cases: the search finds something useful, it finds nothing, and it can’t reach a source at all. The third is where unattended workflows go quietly wrong, so log failures and always show when the last good update ran.
A worked example: a dashboard that updates itself
The Self-Updating Thematic Dashboards cookbook builds one around energy markets and the Iran-Gulf region. On each run the agent uses MCP to pull market benchmarks, news and country information, then rewrites the dashboard from that evidence. Between runs, readers see the last completed update with its date on it.
The code is on GitHub if you want to see how it fits together. It’s a saved example rather than the production version, so treat it as a pattern. To run something similar on your own theme you’ll host the schedule yourself and choose how often it fires.
The most useful exercise happens before any code: decide what belongs on the page. For an energy desk that might be the benchmark prices, the two or three stories explaining a shipping disruption, and an obvious “last updated” stamp. Once that’s written down, the agent’s instructions more or less write themselves.
Keep the receipt next to every number
This is the part that gets you past the committee. Whenever the agent writes a number, store where it came from right beside it: the article title, publisher, publication time, document ID and link, and the passage the agent actually used. Anyone reviewing the note can then open the source and see whether it says what the agent claims.
Two traps catch people. A number needs its context to mean anything: currency, units and the period it covers. And an article published this morning can quote last year’s results, so the article date and the period of the figure belong in separate fields. If the agent calculates something, say a growth rate, keep both inputs and the formula so someone else can redo it.
Record | What to keep |
|---|---|
Reported figure | Value, unit, period, source identifier and the supporting passage |
Calculated figure | The source inputs, the calculation and the rounding rule |
Interpretation | Links to the evidence and the reasoning behind the conclusion |
Missing or old input | Its status, the last observation date and what it affects in the output |
Before anyone relies on the output, open the source links and check them. When two sources disagree, show both and compare their dates and definitions before choosing one. If an input is missing or stale, say so on the page or hold the update back. An old figure should always carry its original date.
What it costs to run
The other thing small funds worry about is the bill. Trying a new data provider, as one person told us, means you “really have to take a leap of faith”. Agent research tends to cost less than people expect, because you pay per token for the ranked passages the agent actually reads, not for whole documents or a seat. The example workflows on our pricing page run from $0.57 for a post-earnings note to $1.82 for a short-thesis memo.
New accounts start with $10 in free credits, and you can see what every token costs as you go.
When building your own server is worth it
Most teams should start with the hosted connection and only build when they hit a real gap. The usual gaps are working from your own company universe, applying your own screen, or searching internal and licensed content in a particular way. Our MCP vs API vs plugins piece explains how the options fit together.
If you do build, the Build Your Own MCP cookbook is a good starting point. It registers a company list using Bigdata entity IDs, screens those companies against a theme and runs several searches at once. The code is public, so you can take only the tools your agent needs.
Someone also has to own it: managing access, shipping updates, handling errors and reading the logs. Give the agent only the permissions the research needs. MCP doesn’t make anything secure by itself; your setup does. For more recurring tasks worth handing to an agent, see An equity analyst’s week.
Key takeaways
MCP connects your AI assistant to research tools. Deciding what to research, when to run and whether the answer holds up is still your job.
Store the source, publication time, reporting period and any calculation next to every number the agent writes, so a reviewer can check it.
Run the workflow by hand before scheduling it, and test three cases: useful results, no results and a source that can’t be reached.
Start with the hosted connection, and build your own MCP server only for a specific gap, such as your own company universe or screen.
Example research workflows on Bigdata.com cost $0.57 to $1.82 a run, and new accounts start with $10 in free credits.
Every small fund that tries AI research hits the same wall. The agent writes a confident paragraph with a revenue figure in it, and someone asks where the figure came from. If the honest answer is “the model said so”, the paragraph is worthless. One fund manager we spoke to put it plainly: “I can’t really present that to my credit committee.”
This post is about getting past that wall. It walks through one workflow, a scheduled update on an investment theme, built so that every number traces back to a document and every stale input is flagged before anyone reads it. The examples use Bigdata.com’s MCP server and two open cookbooks you can copy, but the habits carry over to whatever data you connect.
Pick the job before the tool
Start with something your team already does by hand. For most small funds that’s one of three things: catching up on overnight news across the book, keeping a thematic basket current, or checking whether a story changes a thesis. Bigdata.com has worked examples of portfolio briefings and sentiment scans, plus a custom MCP cookbook for screening companies against a theme. Borrow whichever is closest to your routine.
Then decide who reads the output. A PM catching up at 7am wants a few lines and a link per name. A researcher exploring a theme wants the evidence laid out so they can argue with it. Same data, very different notes, and the agent can’t guess which one you meant. Write down the companies, the sources, the time window and what the finished thing should look like. If you haven’t settled on a connector yet, our guide to financial MCP connectors compares the options.
Write the brief as if a new analyst were picking it up
The instructions do most of the work, and they’re the part people rush. Write them for a capable junior on their first day: which companies, what counts as evidence, what the answer must contain and, the part everyone forgets, what to do when something is missing.
Scheduling is a separate piece. In the dashboard example below, an external automation starts the agent on a timer, and MCP is simply how the agent reaches the data. Keeping those two apart in your head makes debugging much easier later.
A few decisions to make before the first scheduled run:
The cutoff time and timezone. “Daily” means nothing until you say daily as of when.
The search window. A fixed period, such as the last 24 hours, is simplest. “Everything since the last successful run” survives a missed day. A little overlap catches stories that arrive late, as long as you tell the agent what to do with duplicates.
How old is too old. Yesterday’s price and last quarter’s filing are both normal; a price from last week isn’t. Set a freshness rule for each type of input.
Then run it by hand a few times. Before trusting a schedule you want to have seen three cases: the search finds something useful, it finds nothing, and it can’t reach a source at all. The third is where unattended workflows go quietly wrong, so log failures and always show when the last good update ran.
A worked example: a dashboard that updates itself
The Self-Updating Thematic Dashboards cookbook builds one around energy markets and the Iran-Gulf region. On each run the agent uses MCP to pull market benchmarks, news and country information, then rewrites the dashboard from that evidence. Between runs, readers see the last completed update with its date on it.
The code is on GitHub if you want to see how it fits together. It’s a saved example rather than the production version, so treat it as a pattern. To run something similar on your own theme you’ll host the schedule yourself and choose how often it fires.
The most useful exercise happens before any code: decide what belongs on the page. For an energy desk that might be the benchmark prices, the two or three stories explaining a shipping disruption, and an obvious “last updated” stamp. Once that’s written down, the agent’s instructions more or less write themselves.
Keep the receipt next to every number
This is the part that gets you past the committee. Whenever the agent writes a number, store where it came from right beside it: the article title, publisher, publication time, document ID and link, and the passage the agent actually used. Anyone reviewing the note can then open the source and see whether it says what the agent claims.
Two traps catch people. A number needs its context to mean anything: currency, units and the period it covers. And an article published this morning can quote last year’s results, so the article date and the period of the figure belong in separate fields. If the agent calculates something, say a growth rate, keep both inputs and the formula so someone else can redo it.
Record | What to keep |
|---|---|
Reported figure | Value, unit, period, source identifier and the supporting passage |
Calculated figure | The source inputs, the calculation and the rounding rule |
Interpretation | Links to the evidence and the reasoning behind the conclusion |
Missing or old input | Its status, the last observation date and what it affects in the output |
Before anyone relies on the output, open the source links and check them. When two sources disagree, show both and compare their dates and definitions before choosing one. If an input is missing or stale, say so on the page or hold the update back. An old figure should always carry its original date.
What it costs to run
The other thing small funds worry about is the bill. Trying a new data provider, as one person told us, means you “really have to take a leap of faith”. Agent research tends to cost less than people expect, because you pay per token for the ranked passages the agent actually reads, not for whole documents or a seat. The example workflows on our pricing page run from $0.57 for a post-earnings note to $1.82 for a short-thesis memo.
New accounts start with $10 in free credits, and you can see what every token costs as you go.
When building your own server is worth it
Most teams should start with the hosted connection and only build when they hit a real gap. The usual gaps are working from your own company universe, applying your own screen, or searching internal and licensed content in a particular way. Our MCP vs API vs plugins piece explains how the options fit together.
If you do build, the Build Your Own MCP cookbook is a good starting point. It registers a company list using Bigdata entity IDs, screens those companies against a theme and runs several searches at once. The code is public, so you can take only the tools your agent needs.
Someone also has to own it: managing access, shipping updates, handling errors and reading the logs. Give the agent only the permissions the research needs. MCP doesn’t make anything secure by itself; your setup does. For more recurring tasks worth handing to an agent, see An equity analyst’s week.
Frequently asked questions
It depends on how the client, server, authentication and deployment are set up, not on MCP itself. Before connecting, check who can see queries and logs, how credentials are stored and what each tool is allowed to do. The MCP security guidance is a sensible checklist.
Not to connect. The hosted connection takes a few minutes with the setup guide. Running the scheduled dashboard or custom tools does need someone comfortable with code and deployment.
No. The custom cookbook shows how to expose workflows built on Bigdata.com’s REST API as MCP tools. Use the API when your code needs direct control, and MCP when you want an agent to call those tools.
Claude, ChatGPT and Microsoft Copilot all work. If you already use one of them, you can connect Bigdata.com and start asking questions in the same chat window. Developers can also use it in coding tools like Cursor. If your firm manages these apps centrally, IT may need to switch on connections to outside data before you can add it.
Developers
© 2026 Bigdata.com by RavenPack. All rights reserved.
Developers
© 2026 Bigdata.com by RavenPack. All rights reserved.