hi, this is agentconnect. Sandboxes are seen as the standard way to keep agents safe. But problems arise when team runs several agents with different levels of access. That's why we want to solve.
Access to every repo, mcp server, skill, and credential is controlled separately. A2A permissions are set separately in each direction. Delegated tasks run on receiving agent’s side. Only the result comes back. Access is never passed along with it. This keeps permissions isolated and secure.
We have tested it with claude code, codex, and deepseek harness. You can run it with docker compose or deploy it on kubernetes with agent sandbox. Happy to answer questions.
Access to every repo, mcp server, skill, and credential is controlled separately. A2A permissions are set separately in each direction. Delegated tasks run on receiving agent’s side. Only the result comes back. Access is never passed along with it. This keeps permissions isolated and secure.
We have tested it with claude code, codex, and deepseek harness. You can run it with docker compose or deploy it on kubernetes with agent sandbox. Happy to answer questions.