How to Sell Source Code and Workflows to AI Labs (and What They Pay in 2026)

The last generation of coding models learned from public repositories. The next generation is expected to work like an engineer: take a vague ticket, read an unfamiliar codebase, make a change, survive code review, and fix what breaks in production. Public code teaches the syntax. It does not teach the process, because the process was never published. The design argument in a pull request, the incident that led to a refactor, the three failed approaches before the one that shipped: that is the operational history labs are after, and it is one of the sourcing channels mapped in our guide to where AI labs source training data.

The market moved fast on this. In late 2025 Turing launched a program to acquire full private repositories, design docs, testing logs, and internal wikis from defunct startups. Closure platforms now broker code and workspace archives from companies shutting down, and research labs buy them to build realistic environments where agents practice real engineering work. This is the code-shaped corner of a broader shift toward enterprise operational data, and the same logic applies: the value is in how work happened, not in files alone.

A snapshot of source code with no history is worth little to a lab; there is already a lot of code in the world. What commands a price is production history and the context that travels with it.

Notice that half of that table is not code. It is workflow data: the trail of tickets, reviews, discussions, and decisions around the code. If you have ever asked whether you can sell your company's workflows to AI labs, this is the answer. A repository plus the Jira or Linear history plus the engineering chat that references both is a workflow trajectory in raw form, and it is worth considerably more as a connected package than as three separate exports.

There is no universal price per repository, and anyone quoting one without seeing your code is guessing. What the market has published gives you a floor and a shape. One codebase-buying platform lists a 5,000 dollar baseline per repository, adjusted up or down after grading. SimpleClosure, which brokers assets from shuttering startups, has told reporters its deals run from about 10,000 to 100,000 dollars and that it handled roughly a hundred of them in a year. At the top of the range, one founder reported hundreds of thousands of dollars for a company's complete dataset, code included. Those are reported figures from a young market, not a rate card, and most of them describe defunct-company sales rather than licenses from operating businesses.

What moves the number is consistent across buyers: how much production history the code carries, how much surrounding context comes with it (tickets, reviews, docs, chat), how clean the rights are, how unique the domain is (a fintech ledger or a logistics routing engine beats another CRUD app), whether the buyer can run it, and whether the license is exclusive. A single repository is priced as a repository. A whole engineering organization's history, connected across systems and spanning years, is priced as an operational archive, which is a different conversation. The mechanics behind those structures are covered in our guide to how AI data licensing deals work, and the buyer's view of the same factors is in our breakdown of what AI training data costs.

For an operating company, "selling your code" almost always means licensing it: you keep ownership and keep running your product, and you grant a defined right to use the code and its history for defined purposes, for a defined period, exclusively or not. That distinction matters. A non-exclusive license leaves you free to license again and to keep building; an exclusive one commands a premium and closes those doors. Permitted uses matter too: training, evaluation, and environment construction are different rights with different prices. The one case where an outright sale is common is a shutdown, where the company has no further use for the code and the buyer wants a clean transfer. Our guide to selling data to AI companies covers the sell-versus-license reframe in more depth; the same rules apply to code.

Code has rights problems that a video archive does not, and serious buyers check every one of them. You can only license what you own, so the first questions are about ownership: did every contributor, including contractors and agencies, assign their work to the company? Open-source dependencies come next. Permissively licensed libraries are usually fine to leave in place; copyleft code and anything under a license that restricts redistribution needs to be identified and handled, and a dependency inventory is part of any credible package. Then the contents: secrets and credentials in history, customer data in seed files or fixtures, and personal information in tickets and commit messages all have to come out or be excluded from scope. Third-party confidential material, anything under a customer contract, and regulated data stay out. This is the same principle that governs rights-cleared training data in every other category: the buyer's legal exposure depends on your diligence, so the diligence is part of what they are paying for.

Most of the code deals reported so far involve startups that shut down, because a defunct company can transfer everything cleanly and has every reason to. If that is you, package the full history, not just the final branch, and gather the docs and tickets before the SaaS subscriptions lapse and the archives disappear. The step-by-step process from inventory through rights review to deal structure is laid out in our owner's guide to licensing company data for AI.

The larger and less visible opportunity is the operating company with a decade of engineering history and no intention of going anywhere. Those companies license selectively: a retired product line, a legacy system that has been replaced, an internal tool with years of tickets behind it, under a non-exclusive license with clear exclusions. The word that matters is selectively. A process that starts with "give us your whole GitHub org" is not a serious process, and the owner defines what is in scope, what is excluded, and what safeguards apply.

Evaluation datasets and tasksets are a related but different product. Labs pay domain experts to write tasks, grade outputs, and produce reasoning traces, which is a services model rather than a licensing one, and it is the business of the RL environment and evaluation companies. Where an operating company's data enters that market is as the raw material: real tickets with verified resolutions, real support cases with outcomes, and real engineering tasks with the code change that closed them are the inputs environment builders use to make their tasks realistic. If you hold production tasks with recorded outcomes, they belong in the same inventory as your code, and they are priced as workflow data rather than as a finished eval set.

The honest answer is that it depends on history, context, rights, and domain, and the fastest way to get a first read is to inventory what you hold: repositories, years of history, ticketing and docs systems, and how much of it connects. Troveo's free data value assessment takes about five minutes, asks nine questions about your company's systems and history, and returns a score for the AI-training value of what you have, code and engineering workflow included.

Troveo helps companies understand what proprietary data they hold, protect what matters, and selectively license what is valuable. It has paid more than 20 million dollars to rights holders across video, audio, gaming, robotics, and text, and its business data program covers code and structured data alongside communications, documents, and operational systems. The process runs from inventory and rights review through scoping, privacy review, and packaging to buyer matching, with the owner deciding what is in and what is out at every step. Start with the data value assessment, or talk to us about what your engineering history might be worth.