Source code datasets for AI training

Private repositories with commit history, pull requests, code review, tests, CI results and linked issues.

Code and agentsUpdated
SpecificationWhat buyers and sellers agree on
Measured in
Repositories, lines of code, commits, languages and years of history.
Who buys
Coding assistant and coding agent builders, foundation model labs and RL environment vendors.
Where it comes from
Software companies, agencies (for code they own), closed startups and internal tools teams.
Rights to clear
Client ownership of commissioned code, open-source license obligations, contractor IP assignment and acquisition terms.
Privacy
Secrets, keys and credentials scrubbed; customer data in fixtures or logs removed.
Pricing signal
One crypto startup sold its codebase for about $25,000 (MIT Technology Review, 2026). Value depends on history depth and how unusual the domain is.
In short
  • Private codebases with real commit history, tests and reviews are worth more to coding model builders than public code, which most labs already have.
  • MIT Technology Review reported a crypto startup selling its codebase for about $25,000 in 2026, with more pending for workspace data.
  • Secrets, credentials, client-owned code and third-party code under restrictive licenses must be removed first.
  • Buyers value the history around the code, such as pull requests, reviews and linked issues, as much as the code itself.

What coding AI buyers want

Coding models already see vast amounts of public code. What they lack is private, production-grade software with its history: how bugs were found and fixed, how reviewers pushed back, which tests caught what. That history is training data for coding agents.

A reported reference point

MIT Technology Review reported that a crypto startup sold its codebase for about $25,000, with roughly $15,000 more pending for workspace data, through a wind-down platform that has handled more than 150 deals (MIT Technology Review).

Before you offer a codebase

  1. Ownership. Confirm contractor and employee IP assignments, and exclude code built for clients unless you kept the rights.
  2. Open source. Identify vendored third-party code and its licenses; some copyleft obligations travel with the code.
  3. Secrets. Scan full history, not just the current branch, for keys, tokens and passwords, and remove them.
  4. Customer data. Check fixtures, logs, migrations and seed files.
  5. Competition. Licensing core product code may help a buyer build a competing product; decide what you are comfortable with.

Packaging

Buyers prefer a full git history export plus issue tracker and pull request data, with a manifest of languages, frameworks and domains. See also software issue history and agent trajectory data.

Questions

How much is my codebase worth to an AI company?

Reported deals for small startups have been in the tens of thousands of dollars. Long history, good tests, unusual domains and linked issues push value up.

Can an agency sell code it wrote for clients?

Usually not without client consent, because client contracts typically assign the code to the client.

Will buyers run our code?

Often yes, to build training environments. That is another reason to remove secrets from the entire history.

Related guides

Sources

  1. MIT Technology Review: AI labs are buying data from dead startups to train models (October 2026)
  2. Frankfurt Kurnit Klein & Selz: Selling company data for AI training, what to know before you say yes

Need source code datasets, or hold it?

Tell us what you need or what you hold. We check rights and privacy first, then introduce vetted counterparties. You license directly with each other.

  • We reply to every enquiry within two working days.
  • No upfront fees for sellers
  • We never hold your raw data