Copilot vs Cursor vs Claude Code vs Amazon Q: Which Indemnity Covers AGPL Code?

What an indemnity is, and what it is not

An IP indemnity is a contractual promise to defend or compensate the customer if a third party brings a covered claim. It operates after a claim exists. It does not inspect generated code, flag its origin or keep a copyleft-licensed snippet out of a repository.

That distinction matters for procurement. Vendor marketing tends to present indemnity as a form of assurance, and a buyer can read it as a control. The two are different. An indemnity changes who bears the financial consequence of a dispute. A control changes whether the problematic code is accepted into the codebase in the first place. A sound evaluation treats the first as a risk-transfer term to be read closely and the second as a separate requirement to be met by other means.

Indemnity terms for AI coding assistants also differ by product tier. A protection described for one plan may not apply to another plan from the same vendor, so the comparison has to be made per product and per plan, not per brand.

Why AGPL is a harder case than typical copyright claims

Most indemnity language is drafted around copyright infringement: a third party alleges that generated output copies their protected work. AGPL exposure can arise differently. AGPL is a copyleft licence, and its Section 13 is generally understood to extend source-availability obligations to users who interact with a modified version of the program over a network. That is the feature that separates it from the GPL.

For an enterprise, this means the concern is not limited to a snippet being copied. If AGPL-licensed code becomes part of a service that users reach over a network, the question can become whether licence obligations attach to that service. A claim of that kind is about licence compliance, and the remedy a claimant seeks may be compliance or disclosure rather than damages. Many indemnities are written around damages and defence costs for infringement. Whether a given clause reaches a licence-compliance dispute is a matter of its wording.

Whether any specific use triggers AGPL obligations is a legal question for your counsel. This page does not give legal advice.

Evaluation matrix for the four products

Apply each dimension below to GitHub Copilot, Cursor, Claude Code (under Anthropic's commercial terms) and Amazon Q Developer, one plan at a time. Record the document, version and date reviewed for each answer. Treat anything the terms do not state as unaddressed, not as covered.

• Covered claim types: Does the text name copyright infringement only, or does it extend to open source licence-compliance claims such as AGPL or other copyleft obligations?
• Plan or tier eligibility: Which plans carry the indemnity, and is it in the signed agreement or only in a product policy page that can change?
• Activation conditions: Which filters, code-reference settings or public-code controls must be enabled, and who in your organisation can change them?
• Output and input exclusions: Is coverage lost for modified output, user-supplied prompts or context, or custom configurations?
• Procedural requirements: What notice, cooperation and control-of-defence obligations apply, and within what time?
• Network-use claims: Would a claim based on AGPL Section 13 obligations in your own hosted service fall inside the covered claim definition?

How conditions and exclusions narrow the protection

Indemnities for generated code commonly depend on the customer using the product in a particular way. Conditions can include keeping specific filters or code-reference settings enabled, and exclusions can remove coverage where the output was changed or where the claim stems from what the customer supplied. These terms deserve more attention than the headline promise.

Two operational points follow. First, an activation condition is only as reliable as the setting's governance. If individual developers can toggle a required setting, the organisation may lose coverage without anyone noticing. Second, the modified-output exclusion sits awkwardly with normal practice. Developers routinely edit, merge and refactor suggestions, and the edited result may no longer be the unmodified output the clause describes.

Procedural terms matter as well. Notice deadlines, cooperation duties and the vendor's right to control the defence all shape how useful the indemnity is when a claim arrives. Procurement should confirm that internal processes can actually meet them.

Residual AGPL exposure the customer keeps

Even where an indemnity is broad, several exposures remain with the customer. The code is still in the repository. A licence-compliance obligation, if it attaches, may still need to be addressed regardless of who pays for the dispute. Remediation work, disruption to a release and reputational effects are generally not transferred by a financial remedy. Evidence also matters: demonstrating that a setting was enabled and that output was unmodified at the time of generation requires records that exist before a claim, not after.

This is why provenance at the point of generation is a separate requirement from indemnity. NIST's initial public draft GenAI Profile (AI 600-1) includes an action to verify that content provenance complies with relevant laws, including copyright and intellectual property rights, when generating content. It does not address contractual indemnity. The two work together: provenance reduces what reaches the codebase, and indemnity addresses some of what still does.

Trussed provides code provenance covering the origin, licence and security context of code, and enforces policy in the execution path of coding assistants, with actions such as allow, redact and block. Enforced at runtime, not reviewed after the fact, this produces audit-ready evidence of what the assistant was permitted to do.