agentmark

agentmark is a cryptographic provenance tool for AI-generated code. It is designed to show that code passed through a verified autonomous AI pipeline with no direct human write path and to provide a cryptographic audit trail for each commit.

Its verification model combines four proofs in every commit. An output hash checks that the committed code matches the LLM’s raw response byte for byte. A challenge token is embedded in the task and echoed by the model to show that the model processed that specific request. A request ID records a real API call from a cloud provider. An Ed25519 signature from the registered pipeline records pipeline identity, with v1.0 checking the presence and structure of the signature and later public-key verification planned for v1.1. The project also documents failure modes, including cases where a human reproduces the full pipeline, proxy layers that break hash verification, and local models that do not return provider-issued request IDs.

The site lists use cases for agent-only open source projects, enterprise AI governance, regulated industries, EU AI Act Article 50 compliance, and defense supply chain verification. It also says the tool is meant for commit-level traceability and machine-readable provenance in critical systems.

agentmark is available through a Python SDK, a GitHub App, and a GitHub Action. The quickstart shows installation with pip, wrapping an LLM call, signing a manifest, and adding a CI gate for provenance verification. The page names Anthropic and OpenAI support, and the roadmap mentions later support for a public key registry, pipeline registration, worker-side signature verification, a notary service, TEE attestation, and provider-native verification.

It is released under the Apache 2.0 license and is developed by CloudDon Research. The page also describes it as an open standard.