Dropstone

Dropstone carries one memory
across chat, the CLI, and your code.

It applies what it learns in one place to the work you do in another, on one account across chat, the CLI, the SDK, and the editor. Beyond code it places calls, watches an inbox, and monitors a home.

Dropstone Fast 1.7Max
Download Dropstone CLI

Running on models we host ourselves, on infrastructure you can run inside your own walls.

Scroll

Built by an AI research lab advancing every domain of science.

Blankline

About Dropstone

Dropstone began in your terminal as a coding agent. The same agent now places the calls, works the inbox, and keeps an eye on the house.

Dropstone is an AI agent with a memory that follows your account across chat, the CLI, the SDK, and the editor, and it carries what it learns into the real world.

One memory.
Every surface.

Teach me once and it holds, from the CLI to chat to the SDK to the phone call I place on your behalf.

You asked me to watch.

How can I help you today?
Dropstone Heavy 1.7Max

Why Dropstone

Dropstone acts on what it already knows
instead of waiting for a prompt

The difference isn't how clever it is. It's who moves first.

The Challenge

They're very good, and they sit still.

A Tuesday afternoon

Waiting

  • the first moveyours
  • while you're awaynothing
  • what it remembersthis session
  • the weightstrade secret

Close the window and it stops existing.

The coding agents write excellent diffs and stop at the edge of the terminal. The assistants answer well and can't ship one. All of them wait for a prompt, forget you between sessions, and run somewhere you will never see. Nothing happens on a Tuesday afternoon unless you make it happen.

The Solution

I'm already running. You don't have to open me.

A Tuesday afternoon

Running

  • the first movemine
  • while you're awaystill working
  • what it remembersall of it
  • the weightsopen, yours to host

You find out when it's already done.

Proactive by default, from the first session. I watch what you asked me to watch, act on what I already know about you, and hand back the one decision that was actually yours. Same diffs in the terminal, then everything after it, running on open-weight models that are re-evaluated every month.

Measured

Dropstone measures every request,
from the token it spent to the diff it shipped

The point is not the invoice. It is that the effect of the work is visible: what a request used, what it changed, and what it left behind, on the usage page in your dashboard and in the record behind each request.

Memory

Forever

Teach me once and I keep it, across every session, every surface, and every model swap. You never start over.

Context Window

1M

A one-million-token window on every tier, so the agent holds your whole repository in mind at once.

Usage Per Dollar

4x

Up to four times more weekly coding usage per dollar than other AI coding CLIs.

Intelligence Index

60

Kimi K3 behind Dropstone Heavy, the highest-scoring open-weight model on the Artificial Analysis Index v4.1.1.

Data Retention

Zero

Managed inference runs with zero retention upstream, in the United States, on certified providers. Enterprise and commercial data is excluded from training.

Self-learning

Dropstone works out rules
nobody gave it.

It writes the code, places the calls, watches an inbox, and keeps an eye on a house. When the work goes quiet it goes back over how it went: what worked, what did not, and what it should have caught the first time.

That is how it worked out three rules nobody had written down. It got them wrong, noticed, and had them right the next time on work it had never seen. Blankline ran it as a study, then watched the same habit find patterns in mathematics, so it is not only a coding trick.

Read Research