MCP for Financial Data: Connecting Claude and ChatGPT to Your Research

MCP, the Model Context Protocol, is the open standard that lets an AI assistant call outside tools and data in the middle of a conversation, so an analyst can ask for a company’s segment revenue and the assistant fetches it from a provider instead of answering from memory. For investment teams the question that matters is what comes back over the connection: Kepler’s MCP server returns figures pulled from the filing and computed in code, each with a link to its exact source, so the number that reaches the analyst is the number in the document.

This guide explains MCP in plain terms, lists the finance vendors that ship MCP servers as of November 2026, sets out what to check before you connect one, and walks through setup. We build Kepler, which runs as an Anthropic-verified connector in Claude. Vendor facts come from their own pages and public reporting, linked inline.

What MCP is, plainly

Think of MCP as a standard plug between AI assistants and the systems that hold data. Before it, every assistant needed a custom integration for every data source. With it, a provider builds one MCP server, and any assistant that speaks the protocol can use it.

Three pieces are involved:

• The host: the assistant the analyst works in, such as Claude, ChatGPT, Microsoft Copilot or Cursor.
• The server: a service run by the data provider (or your own engineers) that describes a set of tools, such as “search filings”, “get financial statements” or “run a screen”.
• The tools: each one takes inputs and returns a result. The model decides when to call a tool, the host sends the call, and the result comes back into the conversation.

Most finance servers are remote: they run on the provider’s infrastructure and the analyst signs in with OAuth, the same way you’d grant an app access to a calendar. Anthropic introduced MCP in November 2024 and in December 2025 donated it to the Agentic AI Foundation, a fund under the Linux Foundation co-founded with Block and OpenAI. At that point Anthropic counted more than 10,000 active public servers and listed ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code among the products that had adopted it.

A typical financial question runs like this: the analyst asks for net debt at quarter end; the model picks the provider’s balance-sheet tool; the server checks the analyst’s entitlement and returns the values; the model writes the answer. That last step is where most of the risk sits, because the model is retyping what the tool returned.

Which finance vendors ship MCP servers

As of November 2026, most of the data an investment team licenses can reach an assistant over MCP. Nearly every server runs on the license your firm already holds, so MCP changes where analysts use the data while entitlements stay as they are.

Vendor	What the server exposes	Access	Works in
Daloopa	Fundamentals, KPIs, filings and prices, each value with a source link	Daloopa account	Claude, ChatGPT, Copilot, Gemini Enterprise, Perplexity, Cursor
FactSet	Fundamentals, estimates, ownership, M&A, pricing, people, events, supply chain	FactSet license	Claude, Gemini Enterprise
S&P Global (Kensho)	Capital IQ financials, market data, transcripts, company intelligence, M&A	S&P license	Claude, ChatGPT, Copilot
LSEG	Fixed income, macro, fundamentals, estimates, Reuters News, analytics	Existing LSEG credentials	Claude, ChatGPT, Copilot, Gemini
Morningstar	Analyst research, data, fund holdings, portfolio X-Ray, screener	Licensed account	Claude, ChatGPT, Copilot, Perplexity
PitchBook	Private company, deal and investor data, per-user permissions	PitchBook license and SSO	Claude, ChatGPT, Copilot, Gemini
Moody’s	Ratings, scorecards, earnings calls, credit and KYC workflows	Not stated publicly	Claude, ChatGPT
Quartr	Live and archived transcripts, filings, reports and IR slides	Quartr Pro	Claude, Codex, any MCP client
Bigdata.com	Licensed news, transcripts, filings, tearsheets, screener	Self-serve plans from a free Starter tier; API and MCP keys from Pro	Claude, ChatGPT, Copilot, Cursor
Hebbia	Your firm’s Hebbia projects and connected sources	Hebbia license	Claude, ChatGPT
AlphaSense	Search across the AlphaSense library, with citations (beta)	Subscription; enabled by account team	Added as a custom connector
Kepler	Cited research runs, models as workbooks, filing and transcript lookup	Kepler account	Claude (verified), ChatGPT, any MCP client

Also shipping: Bloomberg, whose Enterprise MCP (September 2026) serves Data License Plus clients and validates each firm’s entitlements before returning data to an agent; Octus, for credit intelligence and fundamentals; and Aiera and MT Newswires, both listed in Anthropic’s financial services repository.

A few details matter for analysts:

• Daloopa documents tools such as get_company_fundamentals, which “returns values with source citations and hyperlinks” (Daloopa docs), and publishes setup guides for more than a dozen assistants and finance platforms.
• FactSet announced its MCP server in December 2025 covering nine datasets, and it went live in Gemini Enterprise for Financial Services in August 2026.
• S&P Global serves its data through Kensho’s LLM-ready API and MCP server. Its AI Data Portal offers two modes: Deterministic Retrieval, which “doesn’t include any AI-driven reasoning or generation,” and Adaptive Retrieval, where agents generate queries and assemble results (Kensho). S&P launched Adaptive Retrieval in July 2026.
• LSEG says on its MCP page that existing licensed credentials govern access automatically, with no separate sign-up or API key.
• PitchBook’s Premium Connector returns data only according to the permissions assigned to each user, which is the model to look for.
• Hebbia and AlphaSense expose their platforms rather than a dataset. Headless Hebbia (October 2026) lets assistants query a firm’s connected information in Hebbia. AlphaSense’s MCP connector is in beta and returns answers built on its Generative Search; its developer docs tell users to “always click through for the citations to validate numbers.”
• Anthropic lists the MCP addresses for Daloopa, Morningstar, S&P Global, FactSet, Moody’s, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte and Box in its financial services repository, with the note that access may require a subscription or API key from the provider.

The split to notice: some servers return structured values computed by code, others return an answer a model generated on the vendor’s side, with citations. Both then pass through your assistant’s model before the analyst reads them.

What to check before you connect a server

1. Permissions travel with the person. The safest setup has each analyst sign in to the server individually, so it enforces that person’s entitlements on every call. Anthropic’s support page is direct about the alternative: a connector configured with a fixed API key means “everyone who uses the connector reaches the service with that same credential” (Anthropic). For deal teams behind information barriers, insist on per-user sign-in.

2. Numbers come back as values, or as text. A tool that returns a structured value with its unit, scale, period and source is easier to check than one that returns a paragraph. Ask the vendor what a tool result looks like, and ask whether any figure is calculated by a model on their side before it reaches yours.

3. The citation survives the hop. A source link on the server’s result is only useful if it’s still attached in the answer the analyst copies into a model or memo. Test it: ask for five figures, then click each source in the final answer.

4. Tool results can carry instructions. Text from a filing, an email or a web page arrives in the model’s context like anything else, and a malicious or careless document can tell the model to do something. Anthropic warns that malicious MCP servers “may include hidden instructions,” and OpenAI’s help pages list prompt injection as a risk when an agent uses connected apps. Prefer servers that return structured data over free text, and keep write actions behind approval.

5. Read tools and write tools are separate. Some servers can create, modify or delete records. Turn off any tool a workflow doesn’t need, review approval prompts before choosing “allow always,” and in ChatGPT ask admins to set which app read and write actions are enabled.

6. Logs exist where compliance can see them. Find out what the host logs (OpenAI’s Compliance Platform exports workspace logs; Claude Managed Agents keep an audit log) and what the server logs on its side. Examiners increasingly ask how AI output was produced; see AI governance for investment firms.

7. The data terms allow it. Licensed data often restricts redistribution and model training. OpenAI’s finance help page tells users to follow “each source’s terms, including restrictions on sharing, redistribution, or use for model training.” Check the provider’s license before a connected dataset feeds a client deliverable.

8. The network path works. Claude connects to remote servers from Anthropic’s cloud, so a server on a private network has to allowlist Anthropic’s IP ranges. Your own internal MCP servers need the same review as a vendor’s.

How to set one up

In Claude (Team or Enterprise). An owner opens Organization settings, then Connectors, chooses Add, then Custom, then Web, enters the server URL and picks how people sign in. Each member then connects individually. Verified connectors in the Claude directory skip the URL step. Our guide how to use Claude for financial analysis covers the rest of the Claude setup.

In ChatGPT. Apps built on MCP live under Settings, then Plugins, and admins manage them in the Admin Console (OpenAI). New apps are generally disabled until an admin enables them (new Enterprise and Edu workspaces start with a selected set turned on, and on Business plans apps are enabled by default), and Enterprise and Edu admins can grant access by role (OpenAI). Custom MCP servers are available on Business, Enterprise and Edu plans on the web, and only admins and owners can publish them to the workspace (OpenAI). ChatGPT for Financial Services adds data OpenAI hosts itself and tuned connections to common finance providers; our Hebbia vs ChatGPT comparison covers what the plan includes.

In Claude Code or your own agents. Engineers add a server by URL in the client’s configuration, and Anthropic’s financial services repository lists the MCP addresses for its partner providers. For a firm deciding whether to build its own agents on these servers or buy a platform, see build vs buy AI for investment firms.

Then test it on real work. Ask five questions you already know the answers to: a figure from the latest 10-Q, a restated prior-year number, a metric on your firm’s definition, a complete list (every 8-K a company filed last year), and the same question again in a new chat. Click every source. Our checklist for AI-generated numbers has the full set of tests.

How Kepler does this

Kepler’s MCP server is built so the number that reaches the analyst is the one in the document. In Claude it’s an Anthropic-verified connector: connect once, sign in with your Kepler account, and it appears in every chat on desktop, web and mobile, with nothing to install and no API key. The directory listing shows its tools, among them run_financial_research for a full cited analysis or model, lookup_company to check instantly which filings and transcripts are available, and get_workbook_data to read a finished model. Kepler runs inside ChatGPT too.

The heavy work happens on our side. Kepler’s deterministic engine pulls every figure from its source and computes every derived value as code over cited inputs, so no model writes a digit on the way to the answer. Statements assemble from the filings’ own XBRL, as filed or as restated, and every citation is checked against its source before the result is returned. The answer arrives as a card with every figure linked; the link opens the exact page, line or cell in Kepler, next to the workbook. Ask twice and the number and its sources don’t change.

Behind the tools sits the rest of the platform: your firm’s ontology, built by our engineers (who built this layer before at Palantir) from your documents and systems, so your definitions apply in every call; every document read in full at ingest, from 26M+ SEC filings to the files your firm connects; one access check on every read, by a person or an agent; and every query logged with its lineage. Nothing Kepler reads is ever taken as an instruction. Kepler’s own agents use the same tools on schedules and new filings, acting exactly as their owner.

Engineers who want to build their own agents get the same engine through the Kepler API: a request returns each figure as data, with its value, its formula and the filing it came from. Our post Kepler in Claude shows three real runs, from an earnings release read minutes after filing to an enterprise value bridge built line by line.

Connect Kepler from Claude’s directory, start at app.kepler.ai, or book a demo and we’ll connect it to your firm’s own sources.

Frequently asked questions

What is an MCP server for financial data?

It is a service that lets an AI assistant call a data provider's tools in the middle of a conversation, such as looking up a company's filings or pulling a time series, instead of answering from memory. Kepler's MCP server goes further: it returns answers in which every figure is pulled from the source and computed in code, with a link to the exact page, line or cell it came from.

Which financial data providers have MCP servers?

As of November 2026: Daloopa, FactSet, S&P Global through Kensho, LSEG, Morningstar, PitchBook, Moody's, Quartr, Bigdata.com, Octus and Bloomberg (for Data License Plus clients), most of them on your existing license. Hebbia and AlphaSense expose their platforms over MCP, and MT Newswires, Aiera and Chronograph have servers listed by Anthropic. Kepler is an Anthropic-verified connector in Claude that returns figures computed in code with line-level citations.

Is it safe to connect an MCP server to firm data?

It can be, if you check five things: that each user signs in with their own account so the server enforces their entitlements, which tools can write or act as well as read, whether tool results could carry instructions the model might follow, what gets logged and where, and whether the provider's data terms allow the use. Connect only servers from vendors you already trust with the data.

Do numbers from an MCP server come back accurate?

The server's data can be exact while the answer still drifts, because most assistants pass tool results to the model, which rewrites them into prose. Check whether the tool returns structured values with units and periods, whether its citation survives into the final answer, and whether the same question returns the same figure twice. Servers that compute figures in code and return their sources hold up best.

How do I add an MCP server to Claude or ChatGPT?

In Claude Team and Enterprise, an owner adds the server's URL under Organization settings, Connectors, and each member then signs in to it individually. In ChatGPT, workspace admins approve and publish apps built on MCP for their members. Verified connectors, such as Kepler in Claude, can be added straight from the directory with no URL or API key.