No cloud service, no proprietary format, just files.
Your AI agent's personality and memory can live as plain text files in a folder instead of locked inside one vendor's app. Open that folder in any capable AI tool and the agent just works — same identity, same memory, no re-explaining who it is. That's the whole idea: no cloud service, no proprietary format, just files you can read, copy, back up, and move with a drag.
This isn't a product you install or a service you sign up for. It's a skill — instructions an AI agent reads once, then acts on your folder directly.
One folder = one portable AI team workspace. Three independent paths, use whichever one you need:
- Build — start from an empty folder and stand up a new multi-agent team with portable structure from day one.
- Migrate — already have scattered prompts, a custom-instructions box, or an existing memory setup? Bring it into the skeleton without losing anything.
- Mount — team already built? Connect it to one more AI tool in minutes, no rebuild required.
- Team-level, not another single-persona tool. Built for multiple agents sharing one root and collaborating — not one chatbot's private memory file.
- Plain text, zero dependencies. Everything is markdown plus one JSON config file. No database, no server, no account to create.
- Verified in practice across five harnesses. Claude Code, Codex/ChatGPT Desktop, OpenCode, Copilot (VS Code), and Cursor — each one's actual file-discovery and hook behavior was checked directly, not assumed from documentation.
- Load this skill into your AI tool of choice (however your tool takes skills or custom instructions).
- Tell it which of the three paths you want — build, migrate, or mount.
- Follow along. It will ask what it needs to know (your language, your target folder, what already exists) before writing anything — it won't touch real files without showing you the plan first.
Read this before you rely on anything below the surface:
- v0.4 · docs-first. The reference files (
team-skeleton.md,migration-protocol.md,mounting-guide.md) are the complete, usable specification today — an AI agent reading them can build, migrate, or mount a team right now. - Node-layer scripts are now included (
scripts/). Hook wiring for session-start/stop/precompact plus a concurrency-safe append tool ships with this package; selftest 101/101 and verified on a real team folder (both start and close-out fired correctly). The md layer remains the required foundation; the node layer is optional acceleration. - The node-layer close-out hook is now verified (session-stop fired correctly on a real team folder). What remains unverified is the md-layer self-driven close-out — i.e. an agent on a hookless harness saving its own memory/log purely from AGENTS.md instructions, with no hook to prompt it. Open-session self-drive has production verification; close-out self-drive is designed the same way and expected to behave the same, but hasn't had its own independent verification on record yet.
- New in v0.4 — governance layers for teams that grow large. The skeleton spec now also covers closed-set memory types + a slot-application channel (with an AI-specific "design graveyard" for rejected ideas), a file-lifecycle policy (workspace close-out, backup retention, harness-specific append-only caps), and an optional metadata/relations layer (frontmatter +
links, bound to machine validation). All optional — the minimum viable team is unchanged.
Version 0.4 · Last updated 2026-07-15