rat-stack ships Learn mode: teach your own coding agent via MCP, with adaptive depth and local-only progress
threepointone · x · 2026-10-11
rat-stack launched a "Learn mode" designed to be consumed by your AI agent: paste a prompt, and your agent connects to the ratstack.sh MCP server, reads a public concept deck via learnDeck/learnCard, and learns rat-stack app architecture ideas (effect, alchemy, foldkit on Cloudflare) while doing real work.
Key design details:
- Adaptive depth: new concepts get a full walkthrough (one-sentence claim, TypeScript snippet compiled against pinned versions, text diagram); introduced concepts shrink to a line after seven days; fluent concepts stay silent unless asked.
- Progress and preferences live locally in /.rat-learn/; no skills installed or written to global folders.
- Explicit privacy rules: the agent never sends code, file paths, repo names, or prompts.
- Docs are Markdown by default (Accept: text/html for HTML), naturally agent-friendly.
It's a notable "docs as agent tutorial" pattern: frameworks teaching agents progressively while they build, instead of teaching humans.
More from coding & agent
- Local AI live coding: AMD Strix Halo + OpenCode with an MCP server for requirements — julianharris · 2026-10-11
- Maintainer: good AI PRs come from other maintainers, so we gate AI contributions — vboykis · 2026-10-11
- Stop Overcomplicating AI Agents: It's Just a Loop, Write It Yourself — alex_verem · 2026-10-11
- Dev plans 24/7 agents in VMs powering a cross-device personal AI assistant app — gnukeith · 2026-10-11
- Hugo Bowne-Anderson and Eleanor Berger Launch Agentic Software Factory Course on Maven — intellectronica · 2026-10-11
- Running an agent with its own email inbox: field-tested practices — Jcoulaud · 2026-10-11