Building
Startup data an agent can query. An API and an MCP server, with a source on every fact. There is no dashboard. It opened as a beta on October 13, 2026.
The question is the same one, thirteen years apart. Almost everything around it changed.
| Measure | Mattermark, 2013 | StartupGraph, 2026 |
|---|---|---|
| Who asks | An investor at a browser | An agent with a key |
| The growth signal | Web traffic, social followers, and app downloads | Headcount by quarter, open roles, and GitHub stars, once each has a source |
| How the data got in | Spreadsheets that became crawlers, APIs, and quality checks | Importers, plus agents that submit a fact with its source |
| Corrections | Hundreds of calls and emails after a post | A submitted fact, checked against the page it cites |
| A saved search | An email whenever a new company matched | A screen the agent runs, and a feed of what changed |
| Who built it | Seventy people at the peak | Me, with Claude |
In 2013 ranking a hundred companies by hand took me twenty hours, and the data was stale as soon as I had collected it. Those spreadsheets needed to be code. This time the code also has to show where every number came from.
In the beta
From Companies House, Wikipedia lists, and Y Combinator, with funding rounds for a small set of them and open source projects with their stars.
Search by category, country, stage, headcount, and how recently a company raised. Save a search as a screen and run it again later.
Add a company and the portfolio records its headcount, total raised, and last round. Reading the portfolio later shows what changed since.
Name the investors and get a portfolio that adds their new deals as they appear.
An agent can submit a fact or a whole company along with the page it came from. StartupGraph fetches that page and checks it.
A feed of every fact added or changed since a date, for a list of companies or a whole portfolio.
All of it is reached through the API or the MCP server, with a key. Join the waitlist and watch us build, the launch post on startupgraph.dev.
Agents can add to the data. Nothing they add is trusted until the page they cite says the same thing.
An agent submits a fact or a company with the URL it came from.
The fact goes live at once, marked unverified, and the change is logged.
In the background StartupGraph fetches the source, strips the page, and looks for the company name and the figure, in the forms people write them.
A fact from a known source, or from the company's own site, verifies on that one page. Anything else waits for a second site or a second contributor to agree.
A source that turns out not to exist counts as a strike against the key that cited it. Three strikes and the key is revoked.
A verified fact adds calls to the contributor's key. Anything inconclusive is retried every hour.
The same table is on the front page of startupgraph.dev, and an agent can ask for it.
| Companies | Companies House in the UK, Wikipedia lists, and Y Combinator. |
|---|---|
| Funding rounds | A small set. Nothing after February 2026 is on file yet. |
| Headcount by quarter | Histories exist, but they are not backed by a source yet, so they are served as unverified. |
| Open roles | Not running yet. The weekly pull from public job boards is built and has not had its first run. |
| People | A small set of companies so far. |
| Open source projects | Stars as of the import. |
From the public roadmap, with its labels. None of it has a date.
Building
Next
Later
In 2013 I started Mattermark to answer one question: which startups are growing? The person asking sat at a browser, so we built a dashboard.
You ask Claude which companies in your funds added people this quarter, and Claude asks StartupGraph. Answers are tool results. The agent writes the prose.
Each contributed field comes with the URL it was fetched from, the date, and whether it has verified. It goes in as unverified and stays that way until the source checks out.
Coverage is published per field and per country. An agent that cannot see the gaps fills them with guesses.
2013
After Referly shut down I ranked startups by hand on my blog: Alexa rank, Twitter followers, App Store rank. One list of a hundred companies took twenty hours. We turned the spreadsheets into crawlers and launched Mattermark after six weeks of coding.
June 2017
A post naming Bay Area companies ripe for a Series A, built from a Mattermark screen. The criteria were about fifteen people, a team that had grown over six months, and nine to twenty-four months since the last round. Saved, it emailed me whenever a new company matched.
December 2017
FullContact bought the company. The Mattermark Daily, the newsletter it started from, sent its last issue in May 2018.
December 2025
Companies, headcount over time, funding rounds, and news mentions, for investors, job seekers, and anyone curious. A web page to browse them.
February 2026
Importers for Y Combinator, Wikipedia, SEC filings, GitHub organisations, Hacker News, Companies House, and Product Hunt. The graph reached about 70,000 companies, most of them from the UK register.
July 2026
A restart plan: agents screen, research, and monitor through MCP and the API. Lists, screens, and notes became data an agent writes.
October 12, 2026
The API, the MCP server, and the landing page moved into one Cloudflare Worker. Keys, screens, portfolios, roll-ups, and the fact pipeline went in the same day, built with Claude. The old Laravel app is now only the importer.
October 13, 2026
startupgraph.dev went up with the API, the MCP server, a public roadmap, and a launch post. Keys go out by email, in order.
There is no dashboard, and that is the product decision rather than a gap. Mattermark sold a filter builder. StartupGraph assumes you have an agent. Every call carries a key, so abuse is one key to cut and every key's usage is visible.
In beta at startupgraph.dev. Keys go out by email, in order. The roadmap is a public page and it says what is true today, including that the question the restart was for is not answerable yet.
checked October 13, 2026