PitchBook MCP replaces lookup, not sourcing: a 5-check test
Three vendors now offer to pipe private-company records straight into Claude or ChatGPT, and someone on the exec floor has asked whether that means you can drop a sourcing seat. The question is fair: if the database answers inside the chat window you already use, what is the subscription for? The answer depends on which of three jobs you were actually buying.
Quick answer: PitchBook MCP gives an AI agent live access to a subset of PitchBook's company, deal, investor and fund records, which replaces the lookup work but not the sourcing work. You still need a sourcing tool because an MCP-connected database answers "which companies match these keywords" and cannot answer "which of these is still operating and on my written mandate". The rows it reads carry no recency or relevance signal, so the agent reports stale and off-mandate companies with the same confidence as live ones.
What PitchBook's MCP launch actually shipped
PitchBook announced two things on October 14 at TechCrunch Disrupt: agentic Skills and Agents inside its own platform that turn a natural-language request into a shareable work product, and PitchBook Data MCP, which pipes a subset of its company, deal, investor and fund data into a customer's own tools and models, live on request. Deal Sourcing Skills are promised across 2027, so PitchBook does not consider sourcing solved by the MCP either.
The hosted endpoint sits at premium.mcp.pitchbook.com/mcp, requires SSO plus a seat-based or trial licence, and connects to Claude Code, Codex, Cursor and VS Code over profiles for 8.5M+ private companies, 2.8M deals, 597k investors and 154k funds. PitchBook says Premium Connector usage accelerated nearly 2,000% year to date with tool calls up 267% in six months, so the demand signal is real.
This is no longer distinctive. Crunchbase shipped an MCP server in July and followed on October 6 with Crunchbase AI, a headless intelligence layer installable from ChatGPT's app directory. Grata ships one too, over 22M+ companies with 600 research analysts behind it. Three of the largest private-market databases now deliver through the same pipe into the same window. Records are the commodity. The delivery surface was the differentiator for about ten weeks, and that window has closed.
The three jobs a corp dev team is actually buying
Corporate development teams buy sourcing software for three jobs that get bundled under one invoice and evaluated as one product.
The first is raw records: name, sector, geography, last known round, headcount band. This is what MCP now commoditises, and it is genuinely better in a chat window than in a web UI.
The second is activity verification: whether this company is still trading, still hiring, still shipping, still answering email. No database field says "alive." Records go quiet in two directions and look identical in both. ZoomInfo's own research puts B2B data decay at roughly 2–3% per month, which means a quarterly refresh cadence lets 6–9% of rows drift between updates. Harmonic, a competitor with an obvious interest in the claim, reports user-observed PitchBook record refresh of three to four months.
The third is mandate fit: whether a company belongs on a short list you wrote, with your thesis, your integration constraints and your size band. That judgment lives in a document on your drive, not in anybody's schema.
MCP improves job one and leaves jobs two and three untouched. The honest question is not "database versus platform" but "which of these three am I short of." If your bottleneck is lookup speed, buy the MCP. If it is that half your long list turns out to be dead or irrelevant after two weeks of analyst time, the MCP just makes you faster at generating that same list.
Why a database in your chat window answers confidently and wrongly
An MCP-connected agent cannot flag the absence of a signal, because absence looks like a complete record. A row with a 2022 Series A, no shutdown flag and a tidy website URL reads as a live company to a model, and it has no field to consult that says "nobody has verified this since March."
That failure mode is getting more expensive. Foley reports more venture-backed shutdowns in 2026 than at any point on record, with Q1 2026 closures running roughly 2.6x Q1 2025. PitchBook's own data counts more than 5,000 US startups from the 2019–2022 cohort with no follow-on round, no exit and no confirmed shutdown. Those records say nothing either way, and an agent that reads them reports live matches, because that is the only thing it can do with them.
Affinity's assessment of the connector names related gaps honestly: no pre-formation or pre-visibility signal, no knowledge of your own pipeline history, and no access to proprietary tools like VC Exit Predictor or Manager Scores. Their conclusion points at relationship intelligence, which solves a different problem. Your CRM does not know whether a Hokkaido geothermal developer you have never met is still operating.
There is a governance edge too. Every MCP tool call returns whatever the authorising user can see, uninspected, into a model's context window. Strac's write-up on the setup frames the risk as licensed data leaking outward. The risk I would name first is bad data flowing inward with no audit trail, into a deck that reaches your investment committee.
The 5-check template for testing any MCP-connected source
Run this against your own connection this week rather than trusting a comparison table. Pick one narrow mandate you already know well, ask the agent for companies matching it, take the first ten names, and score them.
- Last-funding recency. Does the agent give you a dated last round? Count how many are more than 24 months old with no explanation. If the agent cannot date the round at all, the row has no recency signal and neither do you.
- Still operating. Independently check whether each company still trades: careers page with a posting from this quarter, a filing, a product release, a LinkedIn post from a named employee. Score it live, quiet or dead. Anything above two dead names in ten is a verification problem, not a coverage problem.
- Record freshness. Ask the agent directly when each record was last updated. If it cannot answer, you have learned the most important thing on this list: your agent's confidence is uncorrelated with its data's age.
- Contactability. Ask who to contact and how. A named person with a role beats a generic info@ address. If every answer is a company website, your outreach is still cold and your conversion maths has not changed.
- The known-good control. Name a company you already know is real, active and squarely on your mandate, and check whether the agent surfaced it unprompted in step one. A source that misses companies you can personally vouch for is telling you about its blind spots for the ones you cannot.
A source that clears all five is worth its seat. One that clears the first and third but fails two, four and five is a records layer, correctly priced and wrongly marketed as a sourcing tool. Chibit sits on the other side of that line: the output is a short list of active, vetted companies matched to your mandate, which is what Innovation Scout returns when you describe a goal rather than a keyword. We wrote up the longer scoring logic in how to score acquisition targets on relevance, activity and fit.
Manual lookup, MCP lookup, and a vetted short list
The three approaches differ less in what they cost than in what they hand back.
| Analyst in a web database | MCP in your chat window | Vetted short list | |
|---|---|---|---|
| Output | A long list, slowly | A long list, in seconds | 10–30 names worth a call |
| Activity check | Analyst does it by hand | Not performed | Done before you see it |
| Mandate fit | Analyst interprets | Keyword proximity | Matched to your written mandate |
| Who to contact | Separate research step | Usually a website | Part of the record |
| Failure mode | Slow | Confident and wrong | Narrower than a directory |
That last cell is the fair criticism. A short list is deliberately smaller than an index, and if your mandate is badly written you will get a confidently narrow answer instead of a confidently stale one. Which is why the mandate document does more work than the tool, and why writing it on one page is the highest-leverage hour you will spend.
Common pitfalls when you add an MCP to your sourcing stack
The most common mistake is treating fluency as verification. The second is dropping the seat you actually needed and keeping the one you did not.
- Measuring the MCP by pipeline volume. Open-innovation practitioners have said this all year: when discovery gets cheap, counting connections makes a team look busier and convert less.
- Assuming the piped subset equals the full product. PitchBook's MCP exposes a subset, and the vendor says so. Proprietary scores and predictions stay inside the platform.
- No named owner for verification. Decide before the first query who checks that a company is alive before it enters a deck. In a chat workflow with no audit trail, "the agent said so" becomes the provenance by default.
- Only testing well-covered geographies. An index-plus-chat-window performs best exactly where coverage is already thick. Run check five on a Daejeon, Hokkaido or Cluj company before you trust it on East Asia or CEE, where the misses concentrate.
Adoption is not the open question any more. Datasite's survey of 1,000 senior dealmakers across 27 countries, run with FT Longitude, found 96% using or exploring AI for sourcing and screening and 62% saying human-only decision-making is no longer defensible. Everyone has the pipe. What separates teams now is whether anyone checks the output.
FAQ
What is PitchBook MCP?
PitchBook Data MCP is a Model Context Protocol server that lets an AI client such as Claude Code, Codex, Cursor or VS Code query a subset of PitchBook's company, deal, investor and fund records directly, instead of searching the PitchBook web platform. It launched on October 14, 2026 and requires SSO plus a seat-based or trial PitchBook licence.
Can PitchBook MCP replace a deal sourcing tool?
PitchBook MCP replaces the lookup layer of a sourcing tool, not the verification or mandate-matching layers. It returns records faster and in the window where you already work. It does not tell you whether a company in those records is still operating or relevant to a specific acquisition mandate.
How do I know whether an AI deal sourcing answer is stale?
Ask the agent for the record's last-updated date and the dated last funding round, then independently confirm the company still trades through a recent job posting, filing or product release. If the agent cannot supply either date, treat every name it returned as unverified rather than live.
Which MCP server is best for corporate development: PitchBook, Crunchbase or Grata?
PitchBook, Crunchbase and Grata expose broadly similar private-company records through the same protocol, so the choice comes down to which index covers your sectors and geographies best and which licence you already hold. None performs activity verification or mandate matching, so pick on coverage and run the five checks above rather than on the MCP feature itself.
Does an MCP-connected database help with off-market targets?
An MCP-connected database performs best on well-covered companies and worst on the off-market and proprietary targets that dealmakers name as their hardest sourcing problem. Companies that have never raised institutional capital, or that sit in thinly covered regions, are precisely the rows that are missing or stale.
Describe your mandate to Innovation Scout and compare its short list against the ten names your MCP returned this morning.
About Andy Chiang
Founder at Chibit
Andy Chiang is the founder of Chibit, a platform that helps corporate innovation, R&D, and M&A teams find active, relevant companies across global innovation ecosystems. He works with buyers who need short lists matched to a real mandate, not directory dumps, with particular focus on green economy, energy, and manufacturing across East Asia, North America, and Eastern Europe. Before Chibit, he spent over a decade in marketing, growth, and go-to-market for technology companies. He writes about operating leverage at Seeking Leverage and hosts Foreign Founders, a podcast and community for immigrant founders, operators, investors, and ecosystem partners. He is based in Brooklyn, New York.
You might also like
Can ChatGPT replace your startup sourcing tool?
Crunchbase AI put a 6M-company index inside ChatGPT. ChatGPT startup sourcing makes retrieval cheap, not diligence better. Run this 5-check test first.
Off-Market Sourcing in Cross-Border M&A: A Practical Sourcing Guide
Learn how corporate M&A teams build proprietary off-market sourcing pipelines in East Asia and Eastern Europe, where platform coverage is thin and cold outreach fails.
Work your Deep Tech Summit Warsaw shortlist before you fly
Five-step playbook for working the Deep Tech Summit Warsaw shortlist before you fly: write the mandate, cut the cohort, verify activity, rank by pilot speed.
Find startups relevant to your goals
Chibit surfaces active, vetted companies matched to your industry and region, so your team starts from a short list worth acting on.
Find Startups