Origin, inside Claude
Ask anything, in plain language, where you already work. Dialectica’s private-markets intelligence answers inside Claude, with no portal to log into and no interface to navigate, every answer cited back into Origin, and the next transcript, snapshot or expert call one step away.
I made this bet before it was strategy and kept it alive as prototypes until customer demand made it urgent. I designed the flows, the model-facing tool surface, the trust model, the answer-to-action loop; I designed, prototyped, and built the MCP app UI. Engineering and data science built and secured production.
Key results
- ≈10× — fewer steps and waits from a question to a cited answer, a step ratio and not a timing, asked in plain language where the analyst already works (Modeled)
- 6 → 1 — sign in, find the company, open the profile, open the transcript, search, copy out: now one question (Modeled)
- Any question, in plain language — across every project you are permitted to see and the whole Origin corpus, right where you already work (Qualitative)
Meet the customer where they are
The bet
For over a year I argued Origin’s next surface wasn’t a better portal but the AI workspace where analysts turn expert calls and transcripts into decisions.
The workflow moved
Research lives in Claude and ChatGPT now; context, collaboration and decks accumulate there.
The data is the business
Expert-backed intelligence is the moat; distribution must never surrender the corpus.
The customer asked for it
One large client told us, in substance, that it would not log into another portal and needed an MCP — and was already sending about half its project requests to a competitor that had one.
The disagreement → the bet
Build the walled garden: make Origin so good customers must come to us.
Meet them where they work: governed, permissioned access inside Claude, with Origin the destination for sources and confirmations.
Concentrating value in the owned platform was a defensible instinct. Customer demand forced the decision, not my argument; I kept us ready. The server is one remote MCP endpoint: Claude today, any MCP client that can reach it next.
Install, sign in, ask
Ask, trace, act
Live in Claude
The replay above runs one session end to end; its caption marks what is live today and what is the designed next release. Below, the anatomy of one answer.
Designing what the model reads
The design work
The MCP has two users: Claude reads the tool contracts, the analyst the evidence and actions. The work lives where no customer looks: permission boundaries, answer anatomy, the friction protecting paid work.
Make the model choose Origin
If the model never picks your tools, nothing else you built gets used.
Exploratory manual trials — my prototype, in-flight build, competitor mock from public tool descriptions; 20-plus runs, memory off, fresh sessions. Claude kept selecting specialised, outcome-described tools over a routed entry point. A signal, not a win rate: I recommended dedicated tools; production shipped ask_origin, origin_get_company, origin_find_experts.
Skills that carry the workflow
Tools return data; the analyst’s job is a workflow the model has to learn.
I authored the Agent Skills in the official Origin plugin under review, and internal skills in daily use at Dialectica. A skill turns tool results into the analyst’s workflow: clarify the ask, search the corpus, offer private transcripts and flag the longer wait, cite every source, end in next steps.
Provenance in every answer
Analysts defend claims to an investment committee; untraceable ones fail.
Answers type their evidence — Dialectica expert, your transcripts, public source — with citations deep-linking into Origin. Expert input is perspective, never ground truth.
Answers that end in action
Research that dead-ends in chat sends customers back to email briefs.
Every answer ends in next steps: unlock a transcript, buy a snapshot, pick experts in the MCP app UI I built. Live today: the deep link lands in Origin, company pre-selected, a click from confirmation. Designed and next: the in-chat expert request, finishing in Origin, brief pre-filled.
The quality bar, written down
A stochastic product breaks silently: spot checks miss regressions, and customers find them first.
I initiated the evaluation strategy and drafted the rubric: seed questions across companies, markets, experts and projects, graded on correctness, citations, completeness, tool choice and next steps, with permission a security requirement, never a graded dimension. Domain operators as ground truth and a calibrated LLM judge come next, not a running system.
In customers’ hands, marketplace under review
Where it stands
As of September 2026: built in-house, pen-tested before release, in daily use at Dialectica and with first customers, the connector and the official plugin submitted to Anthropic’s marketplace and under review. The connector’s public documentation, a page I also designed, is on dialectica.io.
The metric that matters
Project requests started through the MCP against the same client’s email baseline — preference if they shift, growth if they add; no figure until the first cohort produces one.
Coverage and trust
Coverage is not origination-scale yet, and stale Origin data gets contradicted by the host model in front of customers. Work continues on both.
What worked, what I’d change
What worked
Strategy you could install
Entity maps became a Neo4j graph, the graph talk-to-your-data demos, then an MCP around the data-science API that product, engineering and data science ran in their own Claude. Demand set the priority; the installs had already set the direction.
What I’d change
The revenue answered another question
Snapshot revenue proved customers would pay to screen companies they already knew; I read it as demand for finding companies they didn’t. Paying for one job does not validate another. Since then: a year of customer notes in an affinity map, and interviews with the people closest to customers, planned but not yet done.
FAQ
Why an MCP instead of a better portal?
It was never either/or. Origin stays the destination — inspect full profiles, read transcripts, confirm paid requests. The MCP is distribution: the same governed intelligence delivered inside Claude, where the research already happens. The bet is that reach at the decision moment compounds, and a destination alone doesn't.
Isn't proprietary data inside an AI client a leak risk?
That risk framing is exactly why the design leads with governance. Two gates sit in front of every answer: Okta single sign-on over OAuth 2.0, so only provisioned Origin accounts get in at all, and per-user authorization behind it, so the MCP reaches only the projects and transcripts that user already has rights to. It exposes scoped tools that answer from the corpus; it never hands the corpus over. The release was pen-tested. Past those gates the answer lives in the customer’s own Claude workspace, under that customer’s agreement with Anthropic and outside Origin’s control — a boundary the design accepts rather than papers over. Paid work still ends in a confirmation inside Origin, so nothing is bought by accident.
Can a reviewer try it?
Not publicly yet — the connector and the official Origin plugin are under review for Anthropic’s marketplace. Origin’s customers reach it directly today, behind a paid subscription and Okta sign-in. So this page replays the shipped flow, marks the one step that is designed but not yet built, invents figures where the real ones are client-confidential, and ships no open demo: a portfolio must never return paid, permissioned Origin data. The official documentation is public: read it on dialectica.io, a page I designed as well.
Did you build the production MCP yourself?
No. I drove the strategy and kept it alive in prototypes; I designed the flows and the tool surface, designed, prototyped, and built the MCP app UI, authored the skills and drafted the eval rubric; I ran the selection experiments and carried the marketplace submission with legal and compliance: the analysis, the task breakdown, the listing pages and the supporting artifacts. Software engineering and data science designed, built, and secured the production server — architecture, auth, scale. We shipped it together.
What made a designer own the tool descriptions?
Because they're an interface. The model reads a tool contract the way a user scans a screen: naming, scope, and description decide whether your product is ever chosen — before any human sees a pixel. That selection layer is user experience with a new kind of user, so I treated it as design work: prototyped it, tested it adversarially, and handed the pattern to the team.
How do you get to ≈10×?
By counting the two paths to the same answer, not by timing them. Before, on Origin’s own portal: sign in once, then for each company the question touches, find it, open the profile, open the transcript, search it, copy the quote into your notes — five steps a company, and the portal searched one project at a time. Two companies is eleven steps and as many waits. After: one question in Claude, across every project you are permitted to see plus the Origin corpus, answered with citations. One step against eleven is the order of ten. It is a ratio of steps and waits, labelled as a model on the tile and never a stopwatch; the first customer cohort replaces it with a measurement.