The layer, in detail.

One command. It fetches the build for your platform, verifies its checksum, and installs into a directory already on your PATH — no sudo. Needed even for the hosted server: your first key has to be made locally, and nothing on the server could create one for you.

In the project directory you want Myelin to serve. One question decides everything that follows:

This machine only (the default, and the better option for one laptop): a brain is created in ./.myelin/, your host is wired over stdio, and you are done. No account, no network — the key never exists anywhere else.

Hosted server: paste your invite token, pick or create a brain, and the wizard does the rest — the key is generated here and stored in your OS keychain, a connector token scoped to that one brain is issued, your host is wired, and setup ends with a live round-trip against the real endpoint. It finishes with proof, not hope.

Restart your host. It should list Myelin's eighteen tools, four resources, and the myelin-loop prompt — Claude Code shows it under /mcp. Any MCP host works; the wizard prints the three headers for hosts it cannot wire itself.

Use the host the way you already do. Setup has written a Myelin section into your standing-instructions file, so the assistant knows to call recall and match_process before it starts, and to record what it learns as it goes.

Worth doing deliberately on day one: state the conventions you re-explain every week. Myelin does not read your conversation — nothing is stored until the assistant calls a tool — and phrasing decides what survives that call. "We squash before merging, from now on" passes the durable test and can be stored as a standing preference; "squash this one" is refused by design, so a passing remark cannot harden into a permanent rule. Then work normally; after the third success in a task family, or the second time the same correction is logged, a proposal appears:

Point Myelin at a folder of Markdown — an Obsidian vault, a docs directory — and each file becomes one memory, keyed by its path so a second run supersedes a changed note rather than duplicating it. A note renamed in your editor is followed, not duplicated. Anything that looks like a credential is refused, file by file.

That first run links the folder: the path is saved in the brain, and every session from then on re-reads it at startup, so notes you add between sessions arrive on their own. It is a folder you named, read at a moment you can predict — Myelin still watches nothing on its own. myelin serve --vault-sync=false turns it off.

The Markdown export exists so Myelin is never the only thing that can read your brain.

myelin tree lists a directory one path per line — the artifact the project-tree standard reads, so a layout becomes something validate can grade rather than something you argue about. Dependency and version-control directories are excluded by default, and the exclusion list is held to the standard's own by a test. It is a command and not an MCP tool on purpose: both transports share their tool handlers, so a tool that walked a directory would walk the server's disk for anyone who asked.

Every answer the wizard asks for can be a flag, so provisioning a machine is one non-interactive line — including the standing-instructions note, which is written by default on both paths. --auto-invoke takes auto, none, or an explicit list of files. Running your own server is also supported — the full procedure is in deploy/RUNBOOK.md.