# tilde.run

> Tilde turns every AI agent run into a reversible transaction with a versioned composable filesystem, isolated sandboxes, and network-level audit controls for safe production use.

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.*

## Features
- Versioned composable filesystem (GitHub, S3, Google Drive)
- Atomic commit or rollback for every agent run
- Isolated serverless sandboxes per run
- Network egress policy enforcement and audit logging
- Time travel and full commit timeline
- Agent-first RBAC with per-action approval gates
- CLI, Python SDK, and REST API
- Human-in-the-loop approval workflow
- Claude Code integration
- LangGraph support

## Integrations
GitHub, AWS S3, Google Drive, Hugging Face, Claude (Anthropic), LangGraph, Claude Code

## Platforms
WINDOWS, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://tilde.run
- Documentation: https://docs.tilde.run/
- Repository: https://github.com/tilderun
- EveryDev.ai: https://www.everydev.ai/tools/tilde-run