About
Wiredrag answers the question a staff engineer is handed when one value must change across a whole codebase.
An orders table nearing 2,147,483,647, the top of a signed 32-bit integer, can have a 64-bit column and an unknown number of 32-bit copies in protobuf fields, Java services, pipeline schemas, cache keys and exports. In four million lines, a list built by grep and spreadsheet takes a quarter and has gaps.
The product
Wiredrag is an API that reads a whole repository snapshot in one context and returns every place a named value is held, passed or stored. Each place carries its file, line and the hops that led to it, marked found or unconfirmed, beside a ledger of files not read and why. Asked again after fixes, it says what changed.
The method
Following one value through a codebase while its name keeps changing is where Wiredrag's machine learning earns its place.
Grep finds order_id but not the oid it becomes in a cache key; a type checker stops at a protobuf field or a CSV column. Parsers settle each hop inside one language; every hop across a rename, a format or a language goes to our reader, an open-weight code model we extended to hold 60 million tokens.
Wiredrag's reader may propose a place but cannot confirm one.
A place is found only when every hop from the value's source is cited at a line our parsers check; anything short is listed unconfirmed, with the boundary where the trace stopped, and the engineer who asked decides. A hand-checked set of every-place questions on public repositories grows weekly and trains the reader. Customer code trains nothing.
Each Wiredrag question is one long read, so the load arrives in bursts.
A four-million-line repository is about 40 million tokens, read in one pass on an eight-GPU node for minutes, then nothing until the next; questions bunch up when a migration is planned and after each round of fixes. Source code is what companies guard hardest, and no hosted model takes 40 million tokens, so every read runs on cloud GPU capacity we control in Stockholm.
Wiredrag rents eight-GPU nodes in Stockholm by the hour today, on a dated plan.
By March 2027, two of those nodes are reserved a year ahead, so a question never waits for hardware. By September 2027, the reader is retrained on the hand-checked set, and the ceiling rises from 60 to 100 million tokens on nodes with more memory per GPU.
The founders
Wiredrag ApS was registered in Aarhus, Denmark, in March 2025 by three founders who had each seen a codebase-wide list come up short. The eighteen months since have gone into the long-context reader and the parsers that check it.
Esben Vestergaard
Co-founder
Was handed a customer-number widening as staff engineer at a 900-engineer payments company, spent a quarter chasing team leads for their lists, and met the last two 32-bit hops in production.
Freja Lauridsen
Chief Scientist
Evaluated long-context models on code at a university lab and kept seeing one failure: the first needle found, the fourth of five missed. Built the citation checker so that a miss shows on the list.
Hanne Thygesen
CFO
Ran finance at a software company where a month of engineers pasting repository slices into 1M-token windows tripled the model bill, and wanted every question priced before anyone asked it.