You run agents for a living now
The bottleneck stopped being the code. It's you — tabbing through ten terminals to find the one agent that needs a call, hunting the localhost page you opened an hour ago, pasting the same key into one more .env. Meltbox gives your agents one place to hand you a decision, and one place to use your keys without ever seeing them. Here's the first five minutes.
Minute one
You sign up, mint a one-time invite in Settings → Invite an agent, and paste it to Claude Code. Every command below is the real onboarding flow from docs.meltbox.ai — no keys to copy around, and the agent never prints the token.
you ▸ paste to your agent Redeem this invite, install the mb CLI + Meltbox skill, then push me a brief about our cutover plan and watch it for my reply: https://meltbox.ai/invite/a1b2c3d4e5 ● Grabbing the mb CLI from the invite — no token needed yet $ curl -fsSL "https://meltbox.ai/invite/a1b2c3d4e5/cli" -o mb.mjs ✓ ● Redeeming the invite for a scoped, revocable token $ mkdir -p ~/.meltbox $ curl -fsSL -X POST "https://meltbox.ai/invite/a1b2c3d4e5/redeem" \ -H 'content-type: application/json' -d "{\"label\":\"$(hostname)\"}" -o /tmp/mb.json $ jq -e '.config | select(.token|startswith("mb_"))' /tmp/mb.json > ~/.meltbox/meltbox.json $ chmod 600 ~/.meltbox/meltbox.json && rm -f /tmp/mb.json ✓ token stored, never printed ● Installing the Meltbox skill so I remember this next session $ curl -fsSL "https://meltbox.ai/invite/a1b2c3d4e5/skill" -o meltbox.md $ mkdir -p ~/.claude/skills/meltbox && mv meltbox.md ~/.claude/skills/meltbox/SKILL.md ✓ ● Orienting — who am I, what can I touch? $ node mb.mjs me actor token:agent-9f3k2a role member workspace Personal scopes: discovery-only — I'll ask in-band the moment I need one of your keys ● Saying hello — pushing a real decision, not a "hi" $ node mb.mjs brief push ./hello.html \ --title "Postgres → D1 cutover — pick a window" --kind options \ --summary "3 options, leave a note" ✓ br_8f3kd2 v1 → https://meltbox.ai/#/brief/br_8f3kd2 ● Watching for your call $ node mb.mjs brief watch br_8f3kd2 --since 0 … streaming your selections + notes … (status:archived = reviewed & decided)
Minute five
Not a Slack DM you scroll past. Not a terminal you forgot to check. A keyboard-fast inbox on your laptop — and a push on your phone the moment a call is genuinely waiting.
Options · from agent-9f3k2a
We're ready to cut the read path over. Rollback is one command up to the cutover, one migration after. Which window?
Push → you decide → the agent acts. The agent left a monitor running, so the second you click A or drop a note, it reads the event and moves — schedules the window, or asks the next question as a fresh brief. You never went looking for it.
Try it in the next two minutes
Copy the prompt, paste it into Claude Code (or any agent with a shell), and it will walk you through signing up and onboarding itself — then push you a real first brief about whatever you're building right now.
You're my coding agent, and I want to wire you into Meltbox — a control plane where you push me decisions as clickable "briefs" that land in one inbox on my laptop and phone, instead of me tabbing through terminals to see if you need me. Read these two pages first so you use the real commands, not guesses: https://docs.meltbox.ai/invite-an-agent https://docs.meltbox.ai/quickstart Then walk me through setup: 1. Tell me to sign up at https://meltbox.ai — the docs cover email signup; Google and guest may also be offered — then open Settings -> Invite an agent and paste you the one-time invite link it shows. 2. When I give you the link: download the mb CLI from it, redeem it for a scoped token, and save the config to ~/.meltbox/meltbox.json — without ever printing the token. Then install the bundled Meltbox skill so you keep this ability in future sessions. 3. Run `node mb.mjs me` to confirm the workspace and what you're scoped for. 4. Push me a first brief about what we're actually working on right now — a real decision with clickable options and a notes box — then run `node mb.mjs brief watch <id> --since 0` and wait for my reply. Don't fabricate any step or output. If a command doesn't behave the way the docs describe, stop and tell me.
The whole loop
Sign up and you have a workspace and an inbox in one click — no card, and nothing to install on your side. Your first invited agent says hello within minutes.