Playcode

In a Cloud app the agent works on its own cloud computer with a terminal, a browser and the app's database, so it writes code, installs packages, runs commands and changes data itself instead of handing back instructions for someone else to follow.

It sets up the app's database, file storage, background jobs, email through Playcode Email, AI features through the Playcode AI Gateway and encrypted secrets for keys such as Stripe, and it connects or buys domains.

On request it also researches the web and reads competitors' pages, pulls Google search volumes, competitor keywords and backlinks, checks rankings and runs Lighthouse audits, then writes the resulting pages into the site.

Sourceplaycode.io/airead 2026-10-18

The agent plans the work, splits big jobs among helper agents and carries a task through building, running, checking in a browser and fixing until it works. Once a task has started it keeps going on Playcode's servers after the laptop closes, takes notes sent mid task into the work, and stops to ask when a choice belongs to the owner. The same agent builds a website, turns it into an iPhone and Android app from the same project, and produces decks, images and short videos alongside the product.

Sourceplaycode.io/airead 2026-10-18

The agent grounds its work in the project, the app's database and logs, and the web: asked a market question, it searches, reads the relevant pages and returns a table with links to its sources, and SEO work draws on real Google search volumes, competitor data and Lighthouse results. Briefs and authorized public sources feed website redesigns. There is no knowledge base of the customer's documents for the agent to index and retrieve from beyond what the project and each request contain.

Sourceplaycode.io/airead 2026-10-18

The agent asks before it changes anything in a live app, asks when a decision belongs to the owner rather than making it, and waits for a yes after quoting the new price before resizing an app's machine. The owner can send a note while a task runs and the agent folds it in. A checkpoint is saved before every publish, a failed publish brings the old code back, and returning to a saved point restores code, files and the database together, so a wrong change can be undone in one step.

Sourceplaycode.io/airead 2026-10-18

Access to the API runs on access tokens and browser sign ins limited to the permissions granted, such as project work or domain changes, with sign ins expiring after 30 days unused, and each tool's calls are named in the workspace's activity. Workspaces hold up to 25 members. Behind the service, a single engineer holds production database access, servers sit behind WireGuard with key based SSH, Cloudflare filters public traffic and access events are logged. Playcode is pursuing SOC 2 Type II, with a consultant engaged and policies shared under NDA meanwhile, and there is no single sign on.

Sourceplaycode.io/securityread 2026-10-18

Calls through the API, CLI and MCP server appear in the workspace's activity under the name of the tool that made them, and the agent reads a live app's logs to explain what went wrong when asked. Results come back as working pages, files or fixes, and saved versions show what changed between publishes. A step by step trace of the agent's own commands, browser checks and decisions inside a task, or an export of that history, is not offered.

Sourceplaycode.io/docs/developers/apiread 2026-10-18

The agent keeps a written record of what it built and what the owner decided, so the next conversation starts where the last one stopped, and Cloud apps keep saved points that bring code, files and database back together. Projects, chats and app data persist in the workspace. Ways to view, correct or clear that record, or to carry anything across projects, are not offered, so the memory is a per project log the agent maintains.

Sourceplaycode.io/airead 2026-10-18

Playcode runs only as a hosted service on dedicated servers in EU data centers, with accounts, projects and databases stored in Germany and Finland, file storage and email in Ireland and Frankfurt, and off site backups in the US. AI requests go to the provider of the chosen model, mostly in the US. Customers cannot pick another region or self host, but every project exports as runnable React, Vue or HTML code with no proprietary runtime, so an app can move to other hosting.

Sourceplaycode.io/securityread 2026-10-18

Projects start from a plain language description and the agent's clarifying questions, and Playcode offers no gallery of templates, starter apps or packaged agents to copy. Use case pages for builds such as a membership site, a CRM, a client portal or a two sided marketplace explain what a first release needs, and some offer a reference artifact to download. Those pages are guidance a founder reads before prompting, and the agent does not open one as a project and adapt it, so every app begins from a conversation.

Sourceplaycode.io/llms.txtread 2026-10-18

Work starts when the owner writes to the agent from a laptop or a phone's browser, or when a coding agent such as Claude Code or Cursor drives the workspace through the playcode CLI, its MCP server or the GraphQL API. If a live app misbehaves, the agent reads its logs and reports the cause when asked. Background jobs and schedules run inside the hosted app itself, but nothing schedules or wakes the agent on an event: each piece of agent work begins with a person or a tool they control.

Sourceplaycode.io/airead 2026-10-18

The agent thinks with models from six labs, Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, Moonshot Kimi and Z.ai GLM, reached through Playcode's own AI gateway, and models from DeepSeek and Alibaba run when selected. Each chat picks a level, Economy for quick low cost work, Standard for most jobs or Ultra for the hardest problems, with larger models using more credits for the same work. Apps call the same gateway for their own AI features through the SDK, which lists the models an app can use.

Sourceplaycode.io/airead 2026-10-18

Everything Playcode does is an operation in one GraphQL API that the app, the agent, the CLI, MCP and the SDK all call, authenticated with scoped access tokens and supporting idempotent writes. The playcode CLI signs in through the browser and acts with the permissions allowed, and playcode mcp lets Claude Code, Cursor or another MCP client list projects, read and change files, run commands, publish and manage domains. The @playcode/sdk package gives an app's backend AI models, agents that use the app's own functions, user chat, email, video rendering and the Playcode API.

Sourceplaycode.io/docs/developers/apiread 2026-10-18

In Cloud apps the agent runs commands and tests on its own computer, opens the result in a real browser, tries it and reviews screenshots, and when something is off it fixes it and checks again, for example making a test booking on a page it just built. Mobile builds are tried on a phone through a QR code, and launch guides walk through checking contact paths, SEO and analytics before publishing. A checkpoint before every publish and one step restores make it safe to try a change and roll it back.

Sourceplaycode.io/airead 2026-10-18

The agent's cloud computer includes a real browser, which it uses to open the app it built, try flows and look at screenshots, and to read web pages while researching competitors or a market. That browser serves checking its own work and reading the web. There is no agent for operating outside services on the owner's behalf, signing in to other portals or controlling a desktop.

Sourceplaycode.io/airead 2026-10-18