What makes a repository valuable to AI labs.
Frontier labs are licensing real-world source code and git history to train coding models. They have already read the public internet. What is left to learn is how software actually gets built: the commits, reviews, bug fixes and tests that only private repositories carry. This is what they look for, and what they pass on.
Rough thresholds, not hard gates. A repo below them is still worth submitting: falling short is not a rejection.
Who buys, and why it pays.
Frontier AI labs training coding models. Every deal runs through our partner under confidentiality; owners never deal with a lab directly, and neither side is disclosed.
Models have already read the public internet. What is left to learn is how real software gets built: the commits, reviews, bug fixes and tests that only private repositories carry.
A dormant client build, an internal tool, a product you moved on from: code you fully own becomes a licensable asset, without giving up ownership or use.
Five things that make a repository strong.
In the order they ask them.
Production reality
Built for a company, a client or a product, with real users. Commercial applications, client systems, internal tools and backend services beat tutorials, templates and experiments.
Completeness
A whole project, not snippets: source with its tests, docs, configuration and dependencies. Context is what teaches a model how software fits together.
Development history
Years of commits, branches, merged pull requests and fixes. History shows the before and after of every change, which is exactly what coding models train on.
Engineering quality
Clean architecture, real tests, CI, consistent practice, meaningful commit messages. A mature production repo beats a freshly generated one holding the same amount of code.
Scarcity
Code the public datasets do not contain: enterprise systems, specialised industries, less common languages, legacy stacks and multi-service architectures. Rarity is the one dimension that cannot be manufactured after the fact.
What drags a repository down.
Forks and mirrors
If the base is public, models have already trained on it; only original work carries value.
Tutorials & boilerplate
Coursework, templates and scaffolding are the most duplicated code on earth.
Generated bulk
Generated files, vendored dependencies and binaries are excluded from the count, so bloat does not move the number.
Thin history
A single contributor and a handful of commits gives a buyer little to learn from, however clean the code.
Third-party weight
A repo that is mostly someone else's open source is someone else's code.
Falling short is not a rejection
Rights decide everything
- You pick the repositories. Exactly the ones you choose are copied for grading; nothing else is touched, and you can revoke access at your provider at any time.
- Scrub first, surgically. Strip secrets, credentials and client identifiers before syncing. Keep the core logic: the source is what is valued, and removing it lowers the number.
- Nothing moves without you. No repository is offered to anyone without your signature, and grading commits you to nothing.
How a repository is
priced, graded and paid.
Size sets the price, quality decides the sale. Grading is free and private, the estimate obliges you to nothing, and you are paid every time the code is licensed.
These are indicative, deliberately conservative estimates of the owner's share, not guaranteed payouts. Final value depends on the sale and on buyer demand.
Count, grade, licence.
Source tokens at head plus the unique tokens across git history. Generated files, vendored dependencies, binaries and third-party code are excluded before anything is priced.
A model reads the code against a fixed rubric; the repository is never built or run. The same scorer is used before and after syncing, so the grade is consistent throughout.
Non-exclusive by default: you keep using and shipping the code, and the same repository can be licensed to more than one lab and paid for each time.
What the grade measures.
The grade tells buyers what they are getting; the price is set by size at the rates above. When a repo grades below the bar, the working copy is upgraded automatically, and you get the full list of what was added.
Architecture & robustness
Clear layering, real error handling, validated input.
Docs & onboarding
A README a new developer can actually run from.
Code cleanliness
Consistent style, no dead code, linted.
Dependency health
Lockfile committed, lean and current dependencies.
Test coverage
Real tests on real behaviour, run in CI.
Security hygiene
No committed secrets, enforced authorization.
History & maintenance
Meaningful commits, releases, an active repo.
CI/CD maturity
Lint, type-check and tests on every pull request.
Raising the number
Terms in plain words
| Ownership | Stays yours; keep using and shipping the code |
| Standard licence | Non-exclusive · can pay repeatedly |
| Exclusivity | Occasional · premium terms · your call |
| Who sees the code | Review team; a buyer after signing, in confidence |
| Withdrawing | Any time before a licence · 30 days' notice after |
| Payment | Within 15 days of buyer funds landing |
| Cost | None · the rates are what you are paid |
Getting a number for your repos.
Request access
Ask for access and you're walked through the setup: one agreement, already countersigned, then you connect GitHub, GitLab or Bitbucket yourself and select exactly the repositories you want assessed.
Your code never passes through an inbox. Access is scoped to the repos you select, nothing else is touched, and you can revoke it at your provider at any time.
- Token counts for source and history, third-party weight excluded
- A grade across all eight dimensions, with the gaps named
- A total at the standing rates, and what would lift it
Find out what your back catalogue is worth.
Grading is free and private, your originals are never written to, and you can revoke access at any time.