Show HN: Md files as B2B AI agent sandbox – no production data needed
publish.obsidian.mdWe've been running B2B and SMB pilots at Toggle and kept hitting the same wall: "we need access to your data" kills momentum immediately — legal, IT, compliance reviews, weeks of delay.
Our workaround: .md files + live browser telemetry. OpenClaw agents reason over .md files natively. Toggle streams real browser sessions into those files automatically — projects, focus patterns, context switches — so the sandbox stays live without touching any production system. A pilot contact gets a working prototype that feels real because it's built on their actual work behavior, not synthetic data.
What surprised us: the UX signal from .md prototypes is good enough to make real product decisions. By the time a customer graduates to full integration, we already know which workflows to automate and which data connections actually matter.
Full technical writeup covers the pipeline, file architecture, and when this approach breaks down: Pipeline: https://buff.ly/muwUJ9k; Architecture: https://buff.ly/NJBATRz; Pilot playbook: https://buff.ly/ZTo5Jvu
Happy to answer questions on the architecture or the approach. Author here. Happy to go deep on any part of this. The part I'd most like to get feedback on from this community: the decision to use .md files over a proper vector store or database. We made that call for prototyping speed and human readability, but I'm curious if anyone has hit the ceiling on this approach earlier than we did. The honest tradeoffs we found: .md works until you need real-time external data or write actions back to production. For pure read/reason workflows in a productivity context, it's held up better than expected. Also happy to discuss the browser telemetry architecture — specifically how we handle session classification without capturing page content.