Gavriel Cohen

Gavriel Cohen is a software engineer, the creator of NanoClaw, and co-founder and CEO of NanoCo. His work centers on letting AI agents perform useful tasks while restricting the data, credentials and systems they can access. NanoClaw combines an agent harness with container isolation and a core that users customize through code changes and skills.[8][6][4]

Cohen and his brother Lazer were building an AI-focused digital marketing agency when they began using Clawdbot, later called OpenClaw, to manage their sales pipeline through WhatsApp. The agent surfaced Obsidian vault data, tracked deals, assigned tasks and sent reminders. Cohen found it valuable enough to want eight more agents, but worried about its access: even on a dedicated Mac mini, its browser was signed into his Chrome profile and social accounts.[8][3]

Cohen chose Claude Code as NanoClaw’s foundation because it was his preferred AI coding tool. He told The Register that building the surrounding functionality took a weekend, emphasizing that it relied heavily on Claude Code’s existing capabilities rather than recreating them. As NanoClaw gained attention, he and Lazer shifted their focus from marketing to the agent platform.[8]

His architectural distinction was isolation per agent, rather than putting an entire broadly connected system into one container. In his example, a WhatsApp-connected agent received only its assigned group’s messages, rather than all WhatsApp data. It could run bash and install tools within its restricted environment. Cohen summarized the scope he wanted: “I didn't need 3,000 integrations. I needed three.”[8][3]

On March 13, 2026, Docker announced NanoClaw’s integration with disposable, MicroVM-based Docker Sandboxes, providing an execution boundary with a scoped workspace. Cohen’s stated position was that securing agents required boundaries and controls rather than confidence in their behavior: “Teams trust agents to take on increasingly complex and valuable work, but securing agents cannot be based on trust,” he said in the announcement.[6]

In his April 23 V2 announcement, Cohen rejected a forced choice between capable and safe agents: “Either the architecture solves it, or I'm not interested in building it.” The release brought together Docker isolation, OneCLI Agent Vault credential storage and human approvals through Vercel Chat SDK in messaging applications. Cohen credited the partnership with Vercel and Guillermo Rauch. These were the mechanisms behind his argument for greater autonomy with controlled access.[2]

The current NanoClaw architecture uses one Node host to route messages and one Docker container per active session, with inbound and outbound SQLite queues crossing the boundary. Sessions within an agent group share its workspace and memory; separate groups do not. By default, outbound HTTPS passes through OneCLI’s Agent Vault, which injects credentials at request time and applies per-agent policies and rate limits. This separates credential handling from the agent’s execution environment.[5]

NanoClaw’s customization model keeps integrations out of the core until a user chooses them: skills install channel or provider modules into the user’s fork, and coding agents can modify that fork for individual needs. The repository states the contribution principle as “Don't add features. Add skills.” Cohen also committed to keeping NanoClaw open source and developing it as a foundation on which others could build products and businesses.[4][8]

Before the agency and NanoClaw, Cohen worked as a developer and team lead at Wix. He holds a BSc in Physics and Computer Science from Tel Aviv University.[7]