Anthropic licenses Opaque, leaving Confidential MCP without a steward

A developer who built a product on Opaque's Confidential MCP now has a different maintainer upstream. The news is public, the next step for the code is not.

The New Stack reports that Anthropic has struck a licensing and acquihire agreement with Opaque, the confidential computing startup that began as a research project at UC Berkeley's RISELab 10 years ago. The outlet cites The Information for the shape of the deal: Anthropic is licensing Opaque's technology and bringing on some of its engineers, including CEO Aaron Fulkerson and CTO Rishabh Poddar. Opaque's own announcement calls the agreement a "next chapter" that "carries Opaque's work forward" and names no specific commitments to its open-source projects.

What Opaque built that developers depend on

Opaque names the Agent Governance Toolkit, TRACE, Confidential MCP and the Weight Custody Manifest in its own announcement, and lists AgenTrust and the Confidential Computing Summit in earlier posts cited by The New Stack. The summit has already been handed to the Linux Foundation. Nothing has been said in public about who maintains the open-source projects after the licensing agreement takes effect.

The New Stack reports that Opaque's Series B closed earlier this year at $24 million against a $300 million valuation, so this is not a distressed sale. Arjun Jaggi, associate director of R&D innovation at Cognizant, tells The New Stack the deal could be good for the technology if Anthropic puts real resources behind it, since an underfunded startup "might never get it there on its own."

The question Heinrich raises about who verifies whom

Michael Heinrich, co-founder and CEO of 0G Labs, tells The New Stack his caution is about the direction of the check. "The whole point of confidential computing is that you don't have to trust the operator. When the tools that produce that proof sit inside a model vendor, the party being checked now owns part of the checking," he says. The purpose of confidential computing is to let a user verify the operator; a tool for that check that lives inside a model vendor changes who is verifying whom.

Heinrich tells the outlet he is less worried about the code itself, since Confidential MCP is MIT-licensed and anyone can continue using it, and he notes that Anthropic has already given MCP to the Linux Foundation's Agentic AI Foundation. Both he and Jaggi advise developers to pin versions and mirror repositories now.

What this changes today

A team shipping a product that calls Confidential MCP or runs AgenTrust is still shipping it. The upstream project list, the governance framework, the summit and the research all sat inside Opaque, which is now inside Anthropic. For a developer whose compliance story names Opaque's tools, the sentence that reads "maintained by Opaque, an independent startup" is now wrong, and the next audit will ask who the maintainer is.

The practical move is to pin the version your product depends on, mirror the repository, and read the next Opaque or Anthropic post for a named continuation plan. If one does not arrive, treat Confidential MCP and AgenTrust as code you own and can fork rather than as a vendor relationship you can rely on.

Source

Opaque, "From Berkeley research to the future of AI" and The New Stack