AI in regulated & security-conscious workflows

AI is only usable in regulated environments if humans stay in control. Rysh is built that way: every sensitive action is approved by a person, and SecretNAT swaps secrets for tokens before a model request leaves your machine. The server ships as containers with its migrations and a config template, and runs on your own infrastructure.

The problem

Autonomous agents that can run commands are powerful — and risky. In regulated or security-conscious settings you need provable human oversight, no leaked secrets, isolation between teams, and the option to keep everything on-prem.

How Rysh helps

Every tool call that touches the system pauses for a human decision — approve once, approve always for that tool, or reject with a reason the agent must respect. Loop detection and last-prompt-wins prevent runaway behaviour. SecretNAT is on by default: secrets in a model request body are swapped for tokens before the request leaves your machine.

The cloud backend (auth, workspaces, sharing, billing) is a Go + PostgreSQL + NATS + nginx stack that deploys via Docker Compose. The server ships as containers with its migrations and a config template, and runs on your own infrastructure. The collaboration server is proprietary.

Key capabilities

• Human-in-the-loop approval on every sensitive tool call
• Loop detection & last-prompt-wins to stop runaway agents
• SecretNAT on by default: secrets swapped for tokens in model requests
• Per-workspace isolation on the NATS backbone
• The server ships as containers with its migrations and a config template, and runs on your own infrastructure. The collaboration server is proprietary.
• Bring-your-own model API key; approvals keep the human in charge

See it in action

Get started

Spawn a fleet of agents from files on disk, each in its own git worktree.