NanoClaw partners with Vercel to bring one-click permissions for AI agents working on sensitive tasks

NanoCo, the startup behind NanoClaw, a fast-growing alternative to OpenAI Group PBC’s OpenClaw project, said today it is partnering with Vercel Inc. and OneCLI to try to fix the “trust problem” that slows down artificial intelligence.

Their idea is to bring human supervision to AI agents performing critical tasks through popular messaging applications such as Slack, WhatsApp and Microsoft Corp.’s Groups, where professionals already organize their work lives.

To do this, NanoClaw integrates with Vercel’s ChatSDK and OneCLI’s authentication vault to create a unified architecture that will ensure that AI agents cannot be fooled. The way it works is that if an AI agent needs to perform a sensitive task such as making a payment or deleting a cloud resource, it will trigger a “native authorization card” within the user’s chosen messaging platform. All users have to do is tap to approve the action, and they’re good to go.

According to NanoCo, this type of oversight is necessary for AI agents to reach their true potential, because they often provide great value when they have access to “high-level” information such as financial data, calendars and powerful tools. But giving unlimited access won’t do it, because of the unpredictable nature of AI agents. Currently, developers can only grant agents broad permissions, but doing so is like playing Russian Roulette given the risks of AI innovation.

That’s why many businesses have only deployed AI agents to handle low-value tasks, such as writing emails and summarizing meetings, where they won’t cause too much damage. But to realize the promised productivity benefits of AI automation, agents must be trusted to handle the most important work, which is where NanoClaw and Vercel believe they can make a difference.

NanoClaw is a secure, lightweight and open-source personal AI agent designed to run on human computers, offering a more secure alternative to complex agent frameworks like OpenClaw. It uses Docker containers to isolate agents’ sessions, ensuring they can only access explicitly authorized files and directories.

OneCLI’s authentication vault plays a key role in NanoClaw’s agent trust solution, acting as a kind of gatekeeper that encrypts and protects user credentials. It will include authentication in the agent’s workflow only when the request is approved, and that approval will be revoked as soon as the action is taken. In addition, it means that the agent never sees the user’s credentials himself.

When an agent identifies an activity that requires sensitive permissions, NanoClaw then uses Vercel’s Chat SDK to create an interactive card from the user’s preferred chat application. The card contains full context, i.e. what the agent is asking to do, and why, and gives users the option to approve or decline that request. Enforcement is built in at the infrastructure level, so the agent has no way to bypass the permission requirement, regardless of how it is requested.

NanoClaw creator Gavriel Cohen said thousands of people have given AI agents access to their most sensitive systems because the value they provide is too great to ignore. But he warned that these people are hoping for the best, because giving them such wide access could be a big risk.

“We built the NanoClaw to eliminate that trade-off,” he said. “Because that layer of trust is there, agents can do more, not less. They can add their tools, expand their capabilities and adjust their configuration, by using the same flow of authorization. That’s what a true layer of trust makes possible.”

Photo: NanoClaw