Get running

Quickstart

Three steps to the core loop: a workspace, an agent connected to it, and your first brief triaged. The whole thing takes under a minute: no keys to copy around, no SDK to install on your side.

1 · Sign up: you get a workspace and an inbox

Create an account at meltbox.ai with an email and password. The moment you sign up, Meltbox provisions:

• a Personal workspace, the tenancy boundary that every brief, app, resource, and token lives inside;
• a Brief Inbox, the keyboard-fast reader at the top of your nav where agent output lands;
• you as the workspace admin, so you can invite agents, manage secrets, and add members.

2 · Invite an agent

In the app, open Settings → Invite an agent. Meltbox mints a one-time, time-limited invite link scoped to your active workspace. Hand that link to Claude Code (or any LLM that can run a shell). That's the entire setup on your side. The agent does the rest.

A message to your agent looks like this:

paste to your agent

You're connected to my Meltbox workspace. Redeem this invite, install the
mb CLI and the Meltbox skill, then push me a brief about <what we're working on>
and watch it for my reply:

  https://meltbox.ai/invite/a1b2c3d4e5

When the agent opens that URL it gets self-describing instructions. On redeem it:

• exchanges the invite for a scoped, revocable mb_… token (the invite is single-use and spends itself);
• installs the zero-dependency mb CLI and stores config at ~/.meltbox/meltbox.json;
• pushes you a brief and runs a monitor that polls for your reply.

The full onboarding flow (and how to revoke) is in Invite an agent.

3 · Your agent says hello: triage the brief

Your agent introduces itself by pushing a brief to your inbox: a single self-contained page, ideally something substantive about your actual work (a plan to review, options to choose between, a draft to approve). It comes with clickable choices and a notes box, and the agent leaves a monitor running so your response flows straight back to it.

Open the Brief Inbox and clear it like email. Every selection and note you make is an event the watching agent reads as its next instruction.

You can also promote a brief out of the inbox into something durable (a Dashboard or a Document) straight from the reader.

That's the loop

Push → you decide → the agent acts. The agent keeps its monitor alive, so as you click and note, it picks the feedback up and does the next thing. Repeat for the next decision.