tilde.run

Tilde is an agent execution platform that makes autonomous AI code safe to run against real production data. It wraps every agent run in a transaction — compose a filesystem from GitHub, S3, and Google Drive, execute your agent in an isolated sandbox, then commit or roll back atomically. Built by the team behind lakeFS, Tilde brings battle-tested data versioning to the autonomous AI agent era.

• Versioned Composable Filesystem — Mount code from GitHub, training data from S3, and documents from Google Drive as a single ~/sandbox POSIX filesystem; every file is versioned from the first commit and any run can be rolled back instantly.
• Safe Serverless Sandboxes — Each agent run executes in a fresh, isolated container; on clean exit changes commit atomically, on failure nothing changes — no backups to restore or infrastructure to manage.
• Network Isolation & Egress Auditing — Cloud metadata endpoints, private networks, and unauthorized hosts are blocked by default; every outbound request is policy-checked and logged against the agent that made it.
• Time Travel & Audit Trail — Browse the full commit timeline, inspect diffs, and revert any commit instantly; every change is tied to the human, process, or agent that produced it.
• Agent-first RBAC — Agents get their own scoped permissions — never your full user access — with allow, deny, or human-approval policies per agent, per repository, per action, expressed in a simple readable DSL.
• CLI, Python SDK & REST API — Install with a single curl command, run agents from the terminal, orchestrate via the Python SDK, or integrate with the REST API; Claude Code integration lets you spin up sandboxes in plain English.
• Human-in-the-loop Approval Gates — Require human approval before committing sensitive writes; agents pause and surface results for review before any changes land.
• Integrations with popular AI stacks — Works with Hugging Face, Claude, AWS S3, LangGraph, and Google Drive out of the box.