TL;DR: We built Workbench, a VM-based, ephemeral sandbox you (or your AI agents) can spin up, use, and tear down on demand—exposed over the Model Context Protocol (MCP) for code, filesystem, and terminal access. We’re opening a private beta for 25 users; everyone after will be waitlisted. When you join, you’ll get an email with a Discord invite for support and feedback. Issues/examples on GitHub.
Why we built it
Most useful agentic workflows end up writing code. Local tools are great but live on your machine and can be messy for security, reproducibility, and resource contention. We wanted the same “code+tools” experience—but cheap, temporary, and remote—so every run happens in an isolated clean room that auto-expires.
What it is
– Ephemeral VMs: clean, isolated environments for each run
– Full capabilities via MCP: code execution, filesystem, terminal
– Remote & disposable: spin up in seconds; auto-teardown shortly after
– Integrates with MCP-compatible clients (IDEs/assistants) using a simple URL + API key
Usage-based: pay for runtime only (no idle)
Quick start
1. Sign in with GitHub at workbench.brwse.ai.
2. Create an API key.
3. Use the MCP.
Comments URL: https://news.ycombinator.com/item?id=45649397
Points: 1
# Comments: 0
Source: workbench.brwse.ai
