We use analytics cookies to see which pages people read. They are off until you turn them on, and the site works the same either way. The cookies that keep you signed in are essential and always on. Privacy Policy
Case Studies
The star curve is what everyone saw. It is not what we read.
Every chart on this page has three layers. The star curve — the one public signal everyone already watches. A shaded span — the months when one named signal of ours held above the same bar the live product uses: the B2B signal read out of issues, or growth read off the curve's own slope. And a dashed line where the round became public. The composite score never appears on a chart; where signals argue, it only acts.
Four shapes, six cases. Two repositories whose curves said little while their issues filled with Kubernetes deployments, Jira Data Center connectors and self-hosting installs — months before their rounds. Two where the curve itself was the signal, called on the inflection — one of them with a founder signal switching on in the same window. One where no sub-score had crossed a bar, but a person our records already rated was designing the v2.0 architecture in the windows before the seed. And one where the money moved first and the board agreed five weeks later — printed as calibration, because that is what it is.
Or read the six cases first — they are below, with the receipts.
6 monthsHow long the B2B signal held before onyx's seed round became public — while the star curve moved 22%.
31–69%The share of new issues that were enterprise deployments across those six monthly windows. Read, classified, and quotable.
0Decks read, calls taken or introductions used to produce any reading on this page.
Same conclusion
Three shapes the board reads
Each existed as a repository before the money arrived; every round is from our own funding records; every shaded span is the product's own bar applied to reconstructed history, not a story fitted afterwards. Where the signal preceded the round, the chart shows it. Where it did not, the chart shows that too.
onyx-dot-app/onyx
B2B Signal
repository created 2023-04-27 · Seed announced 12 Mar 2025
For six consecutive monthly windows, 31–69% of new issues were enterprise deployments. The star curve moved 22% in the same span.
The curve everyone could see stayed almost flat. The issues did not: for half a year before the round, companies kept filing Kubernetes deployments, Jira Data Center connectors and department-level access controls. That is what the B2B signal reads — and it was firing the whole time.
Receipts — verbatim, from public issues inside the window
“We use the Curators group extensively for different departments to have and manage their own Assistants.”
— issue, Jan 2025
“With my team, we are testing the integration with Jira and the indexing of Jira issues from a specific Jira project. We have Jira hosted on our own (Jira Data Center, not Jira Cloud).”
— issue, Feb 2025
“We deploy the Onyx service in kubernetes and have the application behind a custom ingress that validates authentication with a JWT.”
— issue, Jan 2025
That was one repository.
The same read runs against every repository we track, every day — including the ones whose rounds have not happened yet. Four more cases follow if you want the other two shapes first.
simstudioai/sim
B2B Signal
repository created 2025-01-05 · Series A announced 12 Nov 2025
In the window that closed 121 days before the round, four of nine new issues were teams pulling the product onto their own servers. The demand reading cleared the bar and sim went on the Watch list.
Nine issues came in that month; four were teams standing the product up on their own infrastructure. The visible acceleration came weeks later, and the Series A four months after that. Self-hosting friction is what buying intent looks like in an issue tracker — people do not fight your docker-compose for a toy.
Receipts — verbatim, from public issues inside the window
“A few weeks ago, I pulled the entire repository onto my local server.”
— issue, Jul 2025
“The program requires a self hosted login which blocks the possibility of using it, there is no information in the documents on how to do this or get around it.”
— issue, Jun 2025
“Installed the docker docker-compose.prod.yml and docker-compose.ollama.yml”
— issue, Jun 2025
ComposioHQ/composio
Hyper Growth
repository created 2024-02-23 · Series A announced 22 Jul 2025
14,933 → 23,620 stars in the window the signal fired — +58% in a month, four months before the round.
The opposite shape: here the curve itself is the signal. The momentum reading cleared the same bar the product tags as hyper growth in late February — on the inflection, not after it — and the Series A followed in July.
vm0-ai/vm0
Momentum + founder shift
repository created 2025-11-14 · Seed announced 15 Apr 2026
13 → 956 stars inside the repo’s first eight weeks. In the shaded window momentum sat at its ceiling — and the founder input flipped on, as a builder our records already rated went active. The seed came two months later.
Two months old, vertical curve, and — in the same window — a founder signal that switched on rather than merely existing. One caveat belongs in the record: this tracker is the team’s own build log (the last pre-round window holds 1,880 issues, none of them commercial demand), so the B2B signal correctly read zero throughout. Different engines fire for different companies; this one was called on trajectory and people.
CopilotKit/CopilotKit
Board actionable
repository created 2023-06-19 · Series A announced 5 May 2026
A $27M Series A became public on 5 May. The board first turned actionable 35 days later, twice reached its strongest reading in June, and came back to Watch at the end of July.
The honest direction of this one: the money came first. What the board adds is that it agreed — off public signals alone, five weeks later — with investors who had a data room. Calibration, not clairvoyance.
People our records hold
Atai Barkai @ataibarkai
CopilotKit · Techstars Seattle 23. Previously physics and foundations of quantum mechanics at UPenn.
Alem Tuzlak @AlemTuzlak
Maintains the TanStack AI and devtools packages.
Jordan Ritter @jpr5
Darkridge · Seattle
Michael Vorburger @vorburger
Opened issues in the window before the round — e.g. flagging a deprecated dependency with a published vulnerability in the quickstart install, Apr 2026.
mcp-use/mcp-use
People before the round
repository created 2025-03-28 · Seed announced 12 Feb 2026
Six issues from one engineer our records already rated — an architecture RFC for the v2.0 package split among them — across the two windows before the seed closed. No sub-score had crossed a bar yet; the people reading was what we held.
A different way a repository can be early: not a curve on fire, but a person our records already rated, treating it as infrastructure — designing the v2.0 package split and debugging CSP sandboxing in the two months before the seed. The talent reading was the signal; the round closed the day after the last window.
Receipts — verbatim, from public issues inside the window
“Today, mcp-use is published as a single package that exposes multiple entry points… One package contains multiple environments.”
— RFC issue, Dec 2025
“Zod 4.x uses JIT compilation via new Function()… this triggers Content Security Policy eval violations when widgets run in sandboxed iframes.”
— issue, Jan 2026
People our records hold
Enrico Toniato @tonxxd
Opened six issues in the two windows before the round — the v2.0 package-split RFC, a CSP workaround for sandboxed iframes, and client-completion tests among them.
Method
Where each number comes from
The star curve. Our own direct readings where we have them; below the first direct reading, a monthly reconstruction from the public GitHub event archive. That is why every curve starts near zero — and why points before February 2026 are month-level, not day-level.
The shaded span. Not a story fitted afterwards: we rebuilt each repository's history in monthly windows and asked when one named signal cleared the same bar the live product uses today. The shade is that answer. For the B2B signal, the windows' issues were read and classified one by one; spam and bot floods are excluded from the denominator. One case shades on the people reading instead — a contributor our records rated, filing issues inside the window — and its card says plainly that no score bar had been crossed yet.
Receipts. Verbatim excerpts from public GitHub issues created inside the shaded span, quoted at month precision. Nothing paraphrased.
Scores. Deliberately absent from every chart. The composite score decides actions on the live board; on this page the signals underneath it make the argument. Our full score history was recomputed onto a single current formula, and every historical row carries a stamp saying what evidence backs it.
Why no "we found it on" date. Star, commit and issue history are all available retrospectively, so when a repository happened to enter our board says something about us and nothing about the repository. What matters is the date the score fired, which is what every card shows.
What is missing. A live case — one the board is carrying now, before any round. Candidates exist; they go up when the reading has held long enough to stand behind.
People. People our records hold for the repository — maintainers, and, where we measured it, people who opened issues in the window before the round (the note says which). Only the fields we actually store; where a card lists one person, one is all there is — thin maintainer profiles are common, and closing that is the next thing we are building.
Rounds. From our funding records, verified against the announcing source, and only where the repository existed before the round. Five of the six shaded spans sit before the round tick — four of them by months; the last one closes the day before the seed. The sixth sits after it, and its card says so in the first sentence.
What we are not claiming. That we got there before the money, that we introduced anyone to anyone, or that a high reading causes anything. These are readings our board produced on the dates shown, next to what followed.
Page built from a snapshot taken 31 Jul 2026.
Next questions
Where the six leave you.
Which tracked repositories carry one of these signals today — with no round behind them yet?
How often does a pre-round reading precede the money — and how often does nothing follow?
What does the issue stream look like in the ones that never raise?
One condition runs under every card: a reading is a window, not a verdict. It ends when the next reading drops it — and a shape that contradicts these six goes on this page, not in a drawer.