Show HN: Tilde.run – Agent sandbox with a transactional, versioned filesystem
TL;DR Highlight
Tilde.run, a versioned filesystem for agents, shut down and was folded into lakeFS Mount.
Who Should Read
Backend or platform developers looking for a way to roll back and reproduce file changes made by agents
Core Mechanics
- Tilde.run has been shut down
- The post is titled "We recently shut down tilde.run".
- The team chose to focus on lakeFS instead of continuing it as a separate product.
- State mattered most for running agents
- They say this was the biggest lesson from teams running agents in production over the past year.
- Here, state essentially means the filesystem.
- Object storage alone wasn't enough
- They say agents need fast local access, reproducible state, isolation, and versioning.
- Data, code, config, and intermediate outputs should be composable without constant copying.
- It was folded into lakeFS Mount
- Mount is presented as an agent-native, composable filesystem.
- It combines filesystem ergonomics with versioning, branching, commits, reproducibility, and governance.
- Mount targets a specific workload shape
- It assumes many short-lived environments and large data surfaces.
- It also assumes repeated reads and constantly changing working state.
Community Reactions
- Fatigue with agent sandbox marketing
- One commenter said a new sandbox tool hits the front page almost daily, often with AI-made landing pages, which feels like a bad sign.
- They want the differentiator and use case stated up front.
- Complaint about no open-source or local option
- One commenter had to dig to learn it was a SaaS and said closed-source sandboxes aren't needed.
- They named Apache 2 projects: smolmachines, microsandbox, and boxlite.
- Versioning doesn't cover every risk
- One commenter said version control can't save you if the agent mutates external state.
- Another pointed out that deleting remote GitHub branches or S3 data and data exfiltration need separate controls.
- Experience building something similar
- One commenter built a similar tool but rarely used it because Claude Code was painful to run in a Docker sandbox.
- Another built persistent agent storage themselves with AWS S3 filesystem and Docker.
How to Apply
- Wrap an agent working directory in git
- Pick one task where an agent edits files today.
- Commit before and after the run and confirm you can diff and roll back.
- Write a state requirements checklist
- List the data, config, and intermediate outputs your agent reads and writes.
- Mark whether each needs reproducibility, isolation, or rollback.
- Block external mutations with permissions
- Remove branch-delete and S3-delete rights from the agent's credentials.
- Apply read-only access first, separately from filesystem versioning.
- Evaluate lakeFS Mount for large-scale agents
- This only applies to many short-lived environments reading large data repeatedly.
- Contact lakeFS to check whether your scenario fits.
Try It
# Paste into Claude Code or ChatGPT (simplified version of the article's idea)
Audit the requirements for a versioned filesystem for my agent work.
My agent task: {{what the agent does}}
Data it reads and writes: {{files, DBs, buckets}}
For each item, say whether it's needed and why, in a table: reproducible state, isolation, versioning, branch/merge, rollback.
Finally, list risks that filesystem versioning cannot prevent (external APIs, remote repo deletion, data exfiltration).Terminology
Related Papers
I Am the Reason OpenAI Is Killing Unlimited Chats
Connecting local MCP tools to regular ChatGPT chats gave a parallel coding-agent pool without Codex limits
GPT-5.6 Sol can beat Factorio on peaceful mode
GPT-5.6 Sol in Codex at low thinking finished a peaceful Factorio run with a rocket launch
Do the math: How we took over ChatGPT accounts
One public share link ran attacker JS in a victim's ChatGPT session and could leak private chats
AI isn't replacing SREs, it's deskilling them
Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor
Building an AI-native engineering org
Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%
We ran AI agent swarms in Minecraft and Mindustry
We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.
The Agent SDK
Cursor released Agent SDK, letting you define agents in code and deploy them with one command.