Claude Code contract terms: what to put in an agreement with a developer or agency

A contract for Claude Code work needs the same core terms as any software agreement (scope, price, acceptance, ownership, confidentiality, termination) plus a few that AI-assisted work makes more important: whose Claude account the work runs on, who owns what that account produces, how every AI-written change is reviewed by a person, what data may go into the tool, and what you receive at the end. Leave one of those unstated and you may have nothing in writing to point to if something goes wrong.

This guide lists the terms worth raising with any developer or agency, explains why each matters for Claude Code specifically, and shows what to ask for in plain language. It is written for business owners and managers preparing to sign, so you can have a better conversation with your lawyer, not replace one.

This is not legal advice. Contract law varies by state and country, and the right terms depend on your business, your data and the deal. Have a qualified attorney draft or review any agreement before you sign it.

The clauses at a glance

The scope itself is best agreed in a separate document. Our scope-of-work template covers deliverables, access, review gates and acceptance criteria for a rollout, and the same structure works for a build.

Accounts, seats and tool costs

Claude Code needs a paid Claude plan, an Anthropic Console (API) account or a supported cloud provider such as Amazon Bedrock. Those charges are billed by Anthropic (or the cloud provider), separately from the vendor's fee, so the contract should say whose account is used and who pays. Our guide to Claude Code consultant costs lists the plans and prices as of October 2026.

The account choice is not just about money. Anthropic's legal and compliance page says Commercial Terms apply to Team, Enterprise and Claude API users, and Consumer Terms apply to Free, Pro and Max users. Its data usage page says Anthropic does not train generative models on code or prompts sent to Claude Code under commercial terms unless the customer has chosen to provide its data for model improvement, while for Free, Pro and Max accounts training depends on a user setting. A freelancer on a personal Pro plan and an agency on a Team plan are under different terms.

Options to discuss with counsel:

• The vendor works in seats on your own Team or Enterprise plan, which you control and can remove.
• The vendor uses its own commercial account, and the contract states which plan and what data may be used.
• If a vendor runs Claude Code inside its own product or hosted service for you, Anthropic's legal page sets conditions, including that each end user authenticates with their own credentials and that customers "may not pay for, resell, or intermediate Claude usage on their end users' behalf." Ask the vendor how its setup complies.

Our guide to what happens to your code and data goes deeper on training and retention by plan.

Ownership of the work

Ownership has two layers, and it is easy to think about only one.

The first is between the account holder and Anthropic. Anthropic's Commercial Terms of Service say that, as between the parties and to the extent permitted by law, the customer "retains all rights to its Inputs" and "owns its Outputs," and that Anthropic "hereby assigns to Customer its right, title and interest (if any) in and to Outputs." The Consumer Terms (effective October 8, 2025) contain a similar assignment of Anthropic's rights, if any, in outputs to the user. Note the words "if any": the terms don't decide whether AI-generated material is protected by copyright, which is a separate legal question.

The Commercial Terms also contain an intellectual property indemnity (section K): Anthropic will defend the customer against third-party claims that the customer's paid use of the services, or outputs generated through that authorized use, infringe someone else's intellectual property rights. It has listed exclusions, including claims arising from changes the customer made to the outputs, from combining outputs with technology or content not provided by Anthropic, and from the customer's own inputs. The Consumer Terms contain no equivalent indemnity from Anthropic; under them, the user indemnifies Anthropic. Like ownership, the indemnity runs to Anthropic's customer. If the vendor holds the account, it covers the vendor, not you, so any protection you need has to come from your contract with the vendor. Read section K with your attorney rather than relying on a summary.

The second layer is between you and the vendor. If the vendor's account produced the code, the vendor is Anthropic's customer for that output. Your contract should then say that the deliverables, including code, configuration files, scripts and documentation, belong to you or are assigned to you, usually on payment. Common points to settle:

• What counts as a deliverable. Include the repository, CLAUDE.md files, .claude/settings.json, hooks, skills and any scripts used to make bulk changes.
• Vendor's own materials. Vendors may reuse their own templates, skills and checks. One approach is that they keep ownership and grant you a license to use whatever ends up in your project, for as long as you need it.
• Third-party components. Ask for a list of packages and assets added, with their licenses, so you know your obligations.

The legal status of AI-written code is covered in our guide to who owns AI-written code. Discuss it with counsel if the code is core to your business.

Use of AI tools and human review

If you have your own AI policy, or obligations to your clients or regulators about AI use, the contract should reflect them. Ask the vendor to disclose which AI tools they use on your work and to get your approval before adding new ones that touch your code or data.

More important is review. Anthropic's Consumer Terms state plainly that "Outputs may not always be accurate," and Anthropic's security documentation says users are responsible for reviewing proposed code and commands before approval. Write that into the working rules:

1. Every change is made on a branch in a repository you own.
2. A person on the vendor's side reviews every diff before it reaches you.
3. A check that can fail (a build, a test, a link check or a screenshot) runs before approval.
4. A named person on your side approves each deploy to production.
5. Content facts, prices and claims are approved by you, not by the vendor or the tool.

Our article on AI-generated vs human-edited content explains why the human step matters for content as well as code.

Confidentiality, data and security

A standard confidentiality clause covers what the vendor may disclose. For Claude Code work, add rules on what may be sent to the tool at all, because prompts and file contents go to Anthropic under the account's terms.

• Data categories. State whether customer data, health information or payment data may ever enter a session. For regulated data, Anthropic has specific arrangements (for example, a HIPAA configuration used with a Claude for Enterprise account under a Business Associate Agreement); get counsel involved before any such work.
• Secrets. No passwords or API keys in prompts, chat or CLAUDE.md. Deny rules should block reading files such as .env.
• Permission modes. Anthropic's permission modes documentation says bypassPermissions is for "Isolated containers and VMs only" and "offers no protection against prompt injection or unintended actions." A contract can require that no one runs Claude Code with permission checks bypassed on your systems.
• Access. Staging before production, least-privilege roles, and removal at the end.
• Subcontractors. Whether the vendor may hand work to others, and on whose accounts.

Acceptance, warranties and what not to accept

Acceptance criteria should be things a person can check: "the build passes," "no broken internal links," "redirects return 301," "deny rules block reading env files." Give yourself a review period after each delivery.

For warranties, be wary of promises nobody can keep. No vendor can honestly guarantee bug-free code, a fixed speed-up, cost savings or search rankings, and a contract that leans on such promises gives you little you can enforce. A more useful commitment is a defined period in which the vendor fixes defects against the agreed criteria at no extra charge. Limits of liability, indemnities and insurance are areas for your attorney.

How you pay also shapes the contract. Fixed-scope projects need a tight change-request process; retainers need clear terms on hours, reporting and cancellation. Our comparison of fixed-price vs retainer Claude Code work covers both.

Termination and handover

When an engagement ends, some of what a vendor's Claude Code setup has learned may not be in your repository. Anthropic's docs say auto memory (notes Claude keeps for itself) is machine-local and not shared across machines, and personal files such as CLAUDE.local.md and .claude/settings.local.json stay out of git. The contract should require the vendor to move any rules your project depends on into the committed files before handover.

A handover clause typically covers: all code and configuration committed to your repository, documentation, a list of accounts and access granted, removal or transfer of vendor access, rotation of shared credentials, and a transition period if you want one. Our guide to Claude Code handover documentation lists what you should receive.

Conclusion

A good Claude Code contract settles who owns the account, who owns the work, how every AI-written change is reviewed and approved, what data may enter the tool, and what you receive when it ends. Use this list to prepare, then have an attorney turn it into terms that fit your business; this is not legal advice. If you'd like to see how a vendor handles these points in practice, see our Claude Code development service and ask us the same contract questions you'd ask anyone. More guides are in our Claude Code Guide.

Frequently asked questions