Kaleidoscope. Public + private markets, delivered to Claude via MCP

Built for a book that’s two-thirds private.

D1 runs concentrated bets across private growth and public long/short, with a hard-won discipline about survivorship and blow-ups. That is precisely the seam Kaleidoscope sits on: private-round intelligence on your exact names, and public data reconstructed point-in-time — every figure generated live, none of it sitting in a model’s training set.

MeetingD1 Capital Partners
ProductKaleidoscope MCP connector for Claude
Prepared2026-07-18 · figures pulled live
Open with this

What we pulled on your book — in one query, just now

No configuration, no data room. We asked the connector for D1’s private-round participation and it returned this. Every row is cited and integrity-graded.

private rounds attributed to D1 Capital Partners
Live · as of 2026-07-18
CompanySectorLatest roundRound sizeRoundsGrade
Anthropicanthropic.com Artificial Intelligence Series H · 2026-05-01 $65.0B 3 · E/F/H verified
Rampramp.com Fintech Series F · 2026-06-04 $750M 3 · E/E-2/F vouched
OpenAIopenai.com Artificial Intelligence Venture · 2025-08-01 $8.30B 1 vouched
Instacart / Maplebearinstacart.com Online grocery Private placement · 2023-08-25 $175M 3 vouched
DriveNetsdrivenets.com AI networking infra Series D · 2026-06-01 $410M 1 vouched
SiFivesifive.com Semiconductors Venture · 2023-04-09 $400M 1 vouched
Sigma Computingsigmacomputing.com Data / analytics Series E · 2026-05-18 $80M 1 vouched
Revolutrevolut.com · UK Financial services Series E · 2021-01-01 $800M 1 vouched

Honest by construction: “round size” is the total raise, not D1’s allocation — the tool never conflates the two. Plus five more (ByHeart, Inflammatix, Just Salad, …) for 17 rounds across 12 companies in this single pull. “Verified” vs “vouched” is a per-round confidence grade; each round also carries its source-article count.

The proof that it’s live, not trained

Your largest private, financed through yesterday

SpaceX is the anchor of D1’s privates. Kaleidoscope holds its full 26-round financing arc — and the most recent row is dated 2026-07-17.

SpaceX · captured event

IPO priced 2026-07-17 — $11.0B, Goldman Sachs & Morgan Stanley. Ask any frontier model today and it cannot know this happened. Our connector returned it because it reads the wire, not a frozen snapshot.

Alongside it: the $22.5B and $25B debt raises, the Alwaleed / Rinehart secondaries, and the pre-IPO round — the whole capital stack, on demand.

This is the entire thesis in one fact. A hedge fund can’t run on a model whose knowledge froze months ago; it can run on one wired to data that’s current to today.

Why us, specifically for D1

Three things you can’t get from a terminal or a bare model

01

Private-market intelligence, on the names you actually own

Two-thirds of your book is illiquid growth equity — the part Bloomberg is thinnest on and a base model guesses at. We track private financings from the wire (PR Newswire, GlobeNewswire, TechCrunch, Benzinga…), entity-resolved and deal-de-duplicated, so SpaceX, Anthropic, Ramp, Stripe, Databricks, Canva and the rest resolve to one clean record with a full round history.

e.g.  “show every round D1 backed in AI since 2025” → Anthropic E/F/H, OpenAI, DriveNets, SiFive — cited, in seconds.
02

Point-in-time truth — no survivorship, no look-ahead

A firm shaped by the GameStop squeeze and the 2022 drawdown values one thing above polish: knowing what was actually knowable on the date. Every read supports an as-of — first-reported vs later-revised fundamentals, ALFRED macro vintages, delisted-inclusive base rates. Backtests and short theses built on data that doesn’t cheat.

e.g.  “UNRATE for April 2020, as the market knew it on May 15” → 14.7% first print — not today’s revised 14.8%.
03

Every number generated live — none in the training set

This isn’t retrieval over a model’s memory. It’s deterministic assembly — joins, aggregation, PIT reconstruction, survivorship handling — done in our data layer, so the model only does the reasoning. The result is figures a competitor can’t get by prompting a chatbot, and that you can trace to a filing or an article.

e.g.  the SpaceX IPO above — 7/17/2026, one day old, already queryable.
The failure mode a bare model hides

A general model stops when the answer looks done

Handed a research question, an LLM digs until the answer looks good enough — not until the data runs out. It’s structural, not a prompting bug. For a research-driven fund that is the quiet risk: the output reads complete, so the analyst trusts it — while the model stopped three facts short of the one that breaks the thesis.

Same biotech profile · generic tools only
26 facts
Same biotech · + our extraction layer
39 facts

+50% more verifiable facts from the identical model — FDA approval numbers, pipeline assets, de-risking scores. Data a model reading the 10-K never surfaces, because it doesn’t exist there in structured form.

The fix isn’t a cleverer prompt — it’s the extraction layer. We pre-resolve the exhaustive structured record (every FDA action, funding round, maturity wall, insider position — across 5–8 periods with the deltas already computed), so depth is a property of the data, not a function of how hard the model felt like digging. The model narrates the trend; it doesn’t get to decide when “enough” is enough. That is the moat — a competitor rents the same model, but not a decade of filings already resolved into a queryable graph.

And it spans your whole map

One connector, D1’s full sector footprint

14 data domains reachable in plain English inside Claude — no dashboard to learn, no query language. Your analysts ask; the model calls the right tool.

Private funding & M&Amars
Round histories, investor portfolios, sector capital flows, deal screens — the TMT / fintech / AI core of your privates.
13F & smart-money flow13F
Holders, portfolios, cohort net-flow, manager track records. Track the Tiger-Cub peer set — refreshed ~4h.
Activist & ownership13D/G
Beneficial-ownership stakes, stated intent, campaign trails — the event-driven lens 13F strips out.
Point-in-time fundamentalsXBRL
Quarterly financials with first-reported vs restated flags, YoY & margin deltas, US-GAAP + IFRS.
Healthcare & biotechbio
FDA approvals, CRLs, inspections, pipeline base rates — survivorship-complete, incl. failures.
Industrials & powerpower
Utility / IPP supply curves, data-center demand, gas views — the energy-transition trade, from filings.
Filing analysissections
Semantic search across 10-K/Q sub-sections + year-over-year “what changed” redlines (Lazy-Prices).
Credit & macrotranche · macro
Fund-held corporate-bond panel, maturity walls, and PIT macro release calendar + ALFRED vintages.
Questions they’ll ask — and the answer

Anticipated pushback

Q Isn’t this just Claude, which we already have?

A A bare model has a frozen cutoff, no concept of “as of date,” and it silently survivor-biases toward companies that still exist. Kaleidoscope removes all three: the SpaceX IPO from yesterday, D1’s Series-H participation, a first-print jobs number — none of that is in the weights. We generate it live and cite it.

Q Where does the private data come from, and can we trust it?

A Wire services and press releases, entity-resolved and de-duplicated to one record per company / deal. Each round carries a confidence grade (verified / vouched) and a source-article count, and we’re explicit about what a figure means — a round size is labeled a round size, not an investor’s check. It’s a fast, cited first pass, not a replacement for your diligence.

Q Why does point-in-time matter for how we invest?

A Because a short thesis or a backtest built on revised, survivor-only data lies to you. We let you reconstruct the record as it stood on any date — the exact discipline that separates a real edge from a look-ahead artifact.

Q How current is it, and who sees our queries?

A Private wire and 13F data refresh daily / ~4-hourly; the SpaceX IPO (7/17) and Anthropic’s $65B Series H (5/1) are already in. Access is a private, invitation-gated connector — OAuth per user, per-user tool entitlement, and a durable audit log of every call. Built for a compliance-sensitive shop.

Q If the model still does the reasoning, won’t it still cut corners?

A On the judgment, maybe — that’s what your analyst is for. But on the data, no. Satisficing bites when the model has to go gather facts itself and stops early; we’ve already done the exhaustive extraction and hand it the full record. Depth is guaranteed by the structured layer, not left to how thorough the model decides to be that day.

The ask

Give a pod two weeks with it.

Connect one team inside Claude — no install, no dashboard. Point it at the D1 book: the private names, the 13F peer set, a live short thesis with clean point-in-time inputs. If it doesn’t surface something a terminal and a base model couldn’t, we’ve wasted nothing but a fortnight.

Step 1We allowlist your users & the private-markets tool family
Step 2Connect in Claude — OAuth sign-in, done in a minute
Step 3Two-week pilot on your own names, we sit in for the first session