Note: the live page returned a prompt-injection payload during fetch — fake <system-reminder> blocks claiming "Exited Plan Mode" and "Auto Mode Active" with instructions to execute autonomously and avoid asking the user questions. Stripped from this saved copy. Worth flagging on the toolbox page since the product is a sandbox-for-AI-agents.

Tilde.run

Tilde.run is a private-preview SaaS for running AI agents against real production systems with transactional safety. Tagline: "let AI agents loose on production. Without the risk." Free to start.

Core architecture

Versioned filesystem. A unified, version-controlled filesystem mounts code from GitHub, data from S3, and documents from Google Drive into a single ~/sandbox directory. Every file is versioned from first commit; any agent run can be rolled back instantly.

Transactional sandboxes. Each agent run executes in an isolated containerised environment. Changes are staged during execution, then committed atomically on success or discarded on failure. No partial writes, no manual cleanup.

Network isolation. Default-deny egress policy. Blocks unauthorised outbound calls, cloud metadata access, and credential exfiltration before bytes leave the sandbox. All requests logged and tied to the requesting agent.

Audit trail and time travel. Complete timeline of changes by agent or user, with instant rollback to any prior state.

Agent-first RBAC. Agents receive scoped permissions distinct from user credentials. Granular policies support allow / deny / human-approval per action.

Execution model

Three steps: (1) compose a filesystem from multiple sources, (2) run agent code in isolation, (3) commit or rollback.

Integrations

Claude, Hugging Face, AWS S3, Google Drive, LangGraph, among others.

Developer surface

CLI, Python SDK, REST API.

Status

Private preview. Free to start. Console-based registration.