Project 02 / 05
Ravi
An open-source platform for running AI agents on machines you control, reached from WhatsApp, Slack and the CLI. I help maintain it and work on who may act, where the agent runs and how it is contained.
Receipts
- Mergedfilipexyz/ravi #627+1,801 −60fix(security): stop sending API keys in cleartext; warn when NATS is reachable from the networkKeys stay off the wire, and the bus says so when it is open to more than this machine.
- Mergedfilipexyz/ravi #624+528 −35fix(permissions): explicit-only agents and a stricter bash guardAn agent acts only with an explicit grant, and its shell commands pass a stricter guard.
- In reviewfilipexyz/ravi #662+8,782 −481fix(sandbox): E2B isolation — deny-all egress, API key never in the VM, least-privilege worker, safe signalsSandboxed agents can't reach the network unless allowed, and never see the API key.
Focused on
Technical highlights
- Permissions: agents act only with an explicit grant, behind a stricter bash guard.
- Secrets at rest: owner-only state files, an HTTPS-only gateway, and context keys stored as digests.
- Sandboxes: E2B VMs with deny-all egress, the API key kept out of the VM, and every output treated as hostile.
- Connected sandboxes: a Cloudflare Tunnel and a Gate Worker check short-lived grants before a task reaches a machine.