Setup & integration

Three layers, in order of effort. Step 1 is the only required one — it connects Hilite to your Mental Cache, so captures land in your inbox, not John's. Step 2 turns on the AI features. Step 3 is the optional always-on agent.

Hilite (your Mac)→Mental Cache (your account)→Brain on a VPS (optional)→reports · tasks · Telegram pings

Connect your own Mental Cache

~2 minrequired

Hilite ships with no account attached — the endpoint is prefilled to the Mental Cache API, but capture stays off until you add your own key. The API resolves your key to your user, so everything files to your inbox and nobody else's.

In Mental Cache → Settings → API keys, create a key. It looks like mc_a1b2c3d4…— copy it now; it's shown once.
In Hilite → Settings (menu-bar highlighter icon), paste it into the Mental Cache API key field as:
```
Bearer mc_<your key>
```
Leave the endpoint URL as-is (it's already set to mental-cache.vercel.app/api/external/tasks).
Test: highlight any text, press ⌥Space, hit enter. It should appear in your Mental Cache inbox within a second, with the source app and URL attached.

401 error? The Authorization value must be exactly Bearer mc_… — the word Bearer, a space, then the key.

Turn on the AI features

~2 minpick one

Rewrite-in-place, screen-aware dictation, audio summaries, and screenshot Q&A need a model. Two options; the toggle lives in Settings → AI engine:

Cloud (best quality)

Get a key at console.anthropic.com, paste it into the Anthropic API key field, hit Test key. Costs a fraction of a cent per use. The key stays on your Mac.

Local (free, private)

Install Ollama, then ollama pull gemma4. Flip Run AI locally on and hit Check Ollama. Nothing leaves your Mac; transcription is on-device either way.

Your own brain (always-on agent)

~45 minoptional · advanced

The brain is a tiny VPS that polls your Mental Cache inbox every 10 minutes and acts on what it finds: notes starting with "research…" become published reports, actionable captures get shaped into tasks, "#gl" notes become GitLab-ready issues — with a Telegram ping when something happens, and a morning digest at 7am. Hilite works fine without it; this is the layer that makes captures come back finished.

You need

A small VPS — a $6/mo 2GB DigitalOcean droplet is plenty (Ubuntu).
A Claude subscription (Pro/Max). The brain runs Claude Code headless on your plan — no API bill.
Your mc_ API key from step 1.
Optional: a Telegram bot for notifications (@BotFather, two minutes).

Set it up

SSH in, install Node 20+, then Claude Code:

curl -fsSL https://deb.nodesource.com/setup_22.x | bash - && apt install -y nodejs
npm install -g @anthropic-ai/claude-code

Authenticate on your subscription (creates a long-lived token — do this once):
```
claude setup-token
```

Get the brain bundle from John (scripts + CLAUDE.md — it's not a public repo yet), unpack to /root/brain, and fill in /root/brain/.env:

CLAUDE_CODE_OAUTH_TOKEN=<from claude setup-token>
MENTAL_CACHE_API_KEY=mc_<your key>
TELEGRAM_BOT_TOKEN=<optional>
TELEGRAM_CHAT_ID=<optional>

Add the cron jobs — the poller and the morning digest:

*/10 * * * * flock -n /root/brain/.lock /root/brain/scripts/process-inbox.sh
0 7 * * *   flock -n /root/brain/.lock-digest /root/brain/scripts/morning-digest.sh

Send yourself a test: capture something in Hilite with a note like "research: is X true?" — within ~10 minutes the brain should process it and ping you.

Security rule that matters:the brain box gets exactly three secrets — the Claude token, your Mental Cache key, and the Telegram bot token. Keys that can spend money or read financials (ad accounts, payment APIs) stay on your Mac, never on the VPS. The brain's permissions file scopes it to Mental Cache, web research, and publishing — no shell wildcards.

Debugging: runs log to /root/brain/logs/process-<date>.log. If nothing happens, check the cron is installed (crontab -l) and that .env is loaded — manual runs need set -a && . ./.env && set +a first.

Stuck? Ping John. · Back to Hilite