The vault is neutral ground
Most AI tools want to be the home for your thinking: your knowledge in their database, your annotations in their format, your agent the one they ship, your model the one they lock you to. Then "which AI do I use?" becomes "do I migrate everything?" — and you're fenced inside one vendor's roadmap.
A note.md vault flips it. The vault is a folder of plain markdown with shared conventions — an AGENTS.md that states the house rules, ((file#b-xxxxxx)) block citations for precise references, sidecar .note.md files that hold your judgment, and [[wikilinks]] for a single namespace. Those conventions are a public protocol: any agent can read them, no adapter required. The agents and models become interchangeable workers; the vault is the one thing that doesn't change. It's a git repo, and they're all committing to it.
Assign each job to whoever's best at it
No single agent is best at everything. So don't make one do everything — build a line and put each tool at the station it's strongest at:
| Stage | A good fit | Why |
|---|---|---|
| Overnight automation | OpenClaw / Hermes | Long-running, file-based, self-hosted memory |
| Careful review & revision | Claude Cowork / Code | Strong reasoning; reads your margins before editing |
| Fast drafting & images | ChatGPT (work mode) | Generalist generation, batch image creation |
| In-repo refactors & scripts | Codex | Native AGENTS.md, runs in the working dir |
| Final judgment | You | The one thing no model can generate |
You pick the agent and the model per job — a cheap fast model to triage, a frontier model to reason, a local model for anything private. The vault doesn't care which; it just holds the files they pass between them.
A loop in practice
Here's a real pipeline, four tools and three models over one vault:
- OpenClaw runs overnight, processing a batch of raw notes into
drafts/*.md. - You hand the drafts to Claude Cowork on a careful model — "review and revise these, flag anything shaky."
- ChatGPT batch-generates the hero images into each doc's
_files/folder. - The finished documents land in note.md, where you read them, cut what overclaims, highlight what matters, and leave the notes only you could write.
Four tools, three models, one vault, one orchestrator. Nobody had to share memory or speak a private protocol — they handed off .md files on disk, and your sidecar .note.md annotations were the steering signal for the next one.
Why files make it work
- Anti-lock-in, one layer deeper. Files-over-app frees you from the app; this frees you from the agent and the model. Today's best model is replaced next month — your knowledge shouldn't move with it.
- Collaboration without a platform. Agents hand off
.mdon disk — no shared memory, no private API, no plugin store. One agent's output is the next one's input. - You stay in the loop, at the checkpoint. It's not a black box that runs end to end; it's a line of stations with a human at quality control. Agents write, review, and illustrate — you decide what ships.
- Still just files. No orchestration database, no hidden state. Rules in
AGENTS.md, output in.md, judgment in.note.md— all readable by Obsidian, a CLI, or any agent. Swap out note.md and the vault is still everyone's shared workspace.
Keep the vault in git and every agent write is diffable, attributable, and revertible — an agent's bad day is a git revert, not a tragedy.
Set it up
- Put an
AGENTS.mdat your vault root — grab the conventions from llms-full.txt and add house rules (the hard one: agents never write into*.note.mdsidecars). - Wire up each agent on the same folder: OpenClaw, Cowork, Codex, ChatGPT, Hermes.
- Read and annotate the results in note.md; tell the next agent to read the sidecars first. The loop closes on your disk.
FAQ
Can different AI agents really share one vault?
Yes — that's the design. A vault is plain markdown plus one AGENTS.md describing the conventions. Claude Cowork, Claude Code, Codex, ChatGPT, OpenClaw and Hermes all read and write those files, so you can route each task to whichever agent (and model) is best for it. Keep the vault in git so every write is diffable and revertible.
How do agents hand work off to each other?
Through files. One agent writes markdown into the vault; the next reads it as input. Your annotations live in sidecar .note.md files and act as the steering signal — an agent reads your margins before its next pass. No shared memory or private protocol is needed.
Does this need a special orchestration tool or MCP server?
No. The orchestration is you, and the medium is the filesystem. There's no central database or hidden state — rules in AGENTS.md, output in .md, judgment in .note.md. A Vault MCP server is on the roadmap for agents that prefer a tool interface, but plain files already work today.
Why not just use one AI for everything?
Because no single agent is best at everything. Overnight automation, careful review, fast image generation, private local work, and final judgment are different jobs with different best-fit tools. Splitting them across specialists — over files you own — beats one generalist doing all of it, and keeps you free to swap any worker out.