Troubleshooting

Claude Desktop only reads its list of MCP servers on startup. ModelBrain registers itself the moment it launches, but Claude Desktop won't see that until it's fully restarted — closing just the window isn't enough, the app keeps running underneath on every platform. Make sure ModelBrain's tray icon is actually present first, then fully quit and reopen Claude Desktop: on Windows, open the app's menu (☰) and choose File → Exit, or right-click the Claude icon in the system tray and choose Exit; on macOS, ⌘Q or Claude → Quit Claude from the menu bar. Full steps: connecting Claude Desktop.

Expected. ModelBrain isn't code-signed yet for this release — a paid step being deliberately skipped for now, not a sign anything is wrong. Click More info, then Run anyway. To verify the file yourself first, its SHA-256 is on the download page. Details: installing on Windows.

Only one ModelBrain process can hold your vault open at a time. This usually means two copies of the tray app are running, or a command-line tool (connector_sync sync, a backup) tried to run while the tray already had the vault open. Quit any duplicate ModelBrain instances and try again; a genuinely stuck lock clears itself the moment the process holding it exits, so a full quit-and-reopen of ModelBrain always resolves this.

Indexing runs in the background and a large folder takes a little while on its first pass — give it a few minutes before assuming something's wrong. If it still doesn't show up, check whether the file type or its folder is one of the ones ModelBrain automatically skips (build folders, secret-shaped filenames, unsupported file types) — see what gets read.

ModelBrain picks a reasoning model automatically based on your machine's RAM and CPU cores (see hardware requirements & tiers). Plain search — the default, non-reasoning path — is always fast regardless of tier; reasoning is the slower, deliberate path used only for questions that need facts connected across notes. If reasoning specifically feels slow:

• Run diagnostics to see which tier and model are actually selected.
• A slower machine can be moved to a lighter tier deliberately, trading answer quality for speed.
• The first reasoning question after install or an update downloads the model once (1.5–2.5 GB depending on tier); that download, not reasoning itself, is what's slow the very first time.

A reasoning question can come back saying it isn't available right now instead of an answer. This is a normal, honest response, not an error — here's what each reason means: