Your product. Our computer.
Every agent product is a face and a computer behind it. Pine Computer is the computer, finished:
a whole cloud computer with an agent built in. Bring the model, the logic and the interface, and build the rest on it.
Website · Developer docs · Use cases · Built on Pine Computer · Discord · X
A product built on it, from a recorded call: the video agent talks to the customer while the Computer files the claim.
What it can do · Build it in · Built on it · Plugins · Protocols · Status
| A browser, a desktop, files, a shell, and document and media tools. It works across websites and business software the way a person at a desk does. | Hand over a job, not a script of clicks. Your product keeps its own model, logic and tools, and gets the finished work back. | Sign-ins belong to the Computer. The next job starts where the last one ended, and a job carries on after the chat that started it closes. | It asks before it spends, sends or submits. Watch the live screen, take the controls to type a code or show it how, and hand them back. |
Agents are good at deciding what to do. Most of the world's work still happens behind a sign-in, on a web page, in a file. Pine Computer is the computer for that work, and it is a whole one.
| Does the job | Give it a goal in plain words. It works a real Chrome on a real desktop, across as many tabs as the job needs, with a shell and files beside it. Agent |
| Shows its work | Every step streams back as it happens: what it is doing, each action and how it went, and the screen itself. Agent |
| Asks first | It stops and asks when only a person knows the answer, and before anything that spends, sends or submits. The answer carries the same job on. Steer it mid-job, or cancel. Agent |
| Hands you the controls | Its live desktop sits inside your own page, as a Web Component or a React or Vue component, on a laptop or a phone. A person watches, takes over to sign in or decide, and hands it back. Web SDK |
| Remembers who it is | Sign-ins and passkeys belong to the Computer and come back every time it starts, saved under keys only your server holds. Persistence |
| Learns the job | Turn a job it finished, or one a person showed it, into a skill. You read each skill before any agent uses it, and every later job on that Computer starts out knowing how. Skills |
| Works in parallel | Several sessions on one Computer at once, each with its own window, shell and files, all sharing the same sign-ins. Sessions |
| Makes real files | Fills, edits and merges PDFs; writes Word documents, PowerPoint decks and Excel workbooks; cuts, subtitles and mixes video and audio. Skills |
| Keeps what matters | Artifacts are files the Computer keeps beyond any one session. List, upload and download them even while it is stopped, or share one with a link. Artifacts |
| Keeps secrets | Hand it a password by name. The Computer types it; the model never sees it. Agent |
| Waits without running | Pause a Computer with everything open and carry on later, or stop it and start again from its saved identity. Persistence |
| Works from where you choose | Pick the country its browser reaches the internet from. Overview |
All of it is one SDK away. Your backend creates a Computer and hands it a job; your page shows its screen. The server SDKs are in TypeScript, Go, Ruby and Python, with the same objects and calls in each.
import { Client, Credentials } from "@pinesandbox/computer";
const client = Client.fromEnv();
const credentials = await Credentials.generate();
await store.saveCredentials(credentials.id, credentials.serialize()); // your encrypted database
const computer = await client.createComputer({ credentials });
const session = await computer.createSession();
// Hand over the job. The password is typed by the Computer and never reaches the model.
const goal = "Sign in to the supplier portal as ops@example.com with $PORTAL_PASSWORD and download this month's invoices.";
let turn = await session.agent.run(goal, {
idempotencyKey: message.id,
secrets: { PORTAL_PASSWORD: await vault.get("portal") },
});
// Show the work as it happens.
for await (const event of session.agent.events()) {
const step = event.activity();
if (step) upsertStep(step.id, step.state, step.title); // "Click “Log in”"
if (event.result()) break;
}
// A question ends the turn; the person's reply carries the job on.
turn = await session.agent.wait(turn.turnId);
if (turn.state === "needs_input") turn = await session.agent.answer(turn.turnId, await ask(turn.question));
// Bring the work back, and keep how it was done.
for (const file of turn.result?.artifacts ?? []) await save(file.relativePath, await session.readFile(file.relativePath));
await session.learn(); // a draft skill, waiting for your reviewimport { PineDesktop } from "@pinesandbox/computer-web/react";
// The live screen in your page. Your backend mints a short-lived token for it; no key reaches the browser.
<PineDesktop url={url} token={getDesktopToken} control={driving ? "drive" : "view"} />- Server SDK for TypeScript:
@pinesandbox/computer - Web SDK, the live desktop in your page:
@pinesandbox/computer-web - Every SDK and the full guide: developer docs
The starter is the shortest path from the SDK to a product: a complete app to rename
and ship. A coding agent can build the integration for you with the
pine-computer-sdk skill, which sends it to the guide for the SDK
version you have installed.
The starter: the agent's work on the left, the Computer's live screen on the right.
The SDKs are versioned and released on their own, and their packages are named pinesandbox. This
repository works with the SDK's 0.3 line: @pinesandbox/computer 0.3.5 or newer, @pinesandbox/computer-web
0.3.8 or newer.
A workspace of AI coworkers, a personal agent, a studio, a video agent that files the claim while it talks: each is a face on a computer, and the computer is the hard half. These are whole products with their own model, logic and interface, open source, on the same Computer. Clone one, rename it, ship it.
| Starter | A complete product to rename and ship: sessions side by side, the work live, the controls, skills, a vault, files in and out, schedules. Start here. |
| Roster | A workspace of AI coworkers, each with a Computer of its own: give one a job in its thread, watch its screen beside the chat, approve when it asks. A coding agent built it from the starter in 39 minutes; the log has every step. |
| Tavus video agent | A video agent takes a customer's call and hands the claim to a Pine Computer, which files it while they talk. |
| Drama Studio | A studio for makers of AI vertical short dramas: agents, each on its own Computer, take a screenplay to a cut, and you review where they ask. |
An agent harness decides what should happen and has nowhere for it to happen. Where a harness has a plugin system, Pine Computer goes in as a plugin built on the SDK, using the harness's own credentials and skills.
| DeepSeek Harness | A plugin in its own plugin system, with secrets from its credential store. |
| Claude Code | The tools and the skill together, installed from this repository. |
pine-computer is a CLI, built on the same SDK, that puts one Computer behind the protocols agent apps
already speak. You run it with your own key.
| For | What the app sees | |
|---|---|---|
A tool · pine-computer mcp |
Agents that call tools | Four tools: computer_task, computer_wait, computer_cancel, computer_desktop |
A server · pine-computer serve |
Web apps and agent platforms, over HTTP | The same tools at /mcp, an A2A agent at /a2a with its card, and an AG-UI agent at /agui |
A coworker · pine-computer acp |
Apps that run agents over the Agent Client Protocol | A bot. Each chat thread is a thread on the Computer; its steps show as it works; its summary is the reply |
The terminal · pine-computer run "…" |
You | One job, its steps as it goes |
You need a Pine Computer project key (pinecomputer.io) and Node.js 22 or newer.
git clone https://github.com/PineComputer/pine-computer && cd pine-computer
npm install && npm run build && npm link # puts `pine-computer` on your PATH
mkdir -p ~/.pine-computer && chmod 700 ~/.pine-computer
printf 'PINE_SANDBOX_API_KEY=…\nPINE_SANDBOX_ENDPOINT=…\n' > ~/.pine-computer/env && chmod 600 ~/.pine-computer/env
pine-computer run "Open https://news.ycombinator.com and tell me the top 3 stories."Claude Code, as a plugin:
/plugin marketplace add PineComputer/pine-computer
/plugin install pine-computer@pine-computer
Any agent with an MCP setting:
{ "mcpServers": { "pine-computer": { "command": "pine-computer", "args": ["mcp"] } } }
![]() As a coworker: a bot in the chat, with its own screen. |
![]() As a tool: another agent hands the job over and relays the question. |
Through a protocol a job still streams its steps, asks its question, keeps its thread, shows its live
desktop on a page of its own, brings files back and takes secrets by name. Skills, artifacts, parallel
sessions and the desktop inside your own interface come with the SDK. docs/cli.md is the
full reference, and skills/pine-computer teaches an agent when and how to
hand work over.
What has been run end to end:
| State | |
|---|---|
| Claude Code, over MCP | Run |
| Codex, over MCP | Run |
| OpenMausBot, as a coworker (ACP) and as a tool (MCP) | Run |
A2A agent and hosted MCP (pine-computer serve) |
Run against the protocol's own clients |
CopilotKit OpenBot, over AG-UI (pine-computer serve) |
Run against the protocol's own client; not yet run inside OpenBot |
| Cursor, Claude Desktop, nightly-labs OpenBot | Configuration given, not yet run by us |
| DeepSeek Harness plugin | Loads in dsh, with its tools and skill registered; no job run through it yet |
Apache-2.0. Built by Pine AI.

