Atoms

Adrian connects to Google Ads over OAuth and creates search campaigns, then pauses, enables, deletes or rebudgets them and can apply Google's recommendations, while Atoms can create a GA4 account, property and stream, add the tracking code and set up a Search Console property. Projects can use Supabase with read and write access to its database, auth, storage and edge functions, take Stripe payments, push to GitHub, and read and write Linear, Asana and Todoist tasks, with files from Box, Dropbox and Google Drive.

Sourcehelp.atoms.dev/en/articles/12129582-ads-agentread 2026-11-08

Mike, the team leader agent, splits a request into a plan, routes each part to the product manager, architect, engineer or a specialist agent, and asks questions through cards. Work hands off on its own, from Sarah's SEO pages to Alex for building and from Adrian to Alex for conversion tracking, several agents can take independent tasks from one message with a queue behind them, and Race Mode on Max runs several agent teams on different models in parallel.

Sourcehelp.atoms.dev/en/articles/12129548-working-with-agentsread 2026-11-08

Agents read the project's files and state, attachments of up to 100 MB a file and folders of up to 500 files, references to other projects, connected tools such as Google Drive, Box and Dropbox, and web pages through Deep Research with citations. Atoms offers no knowledge base or document library that persists across projects for agents to search, so grounding comes from what each project holds and what a person attaches.

Sourcehelp.atoms.dev/en/articles/12129597-file-types-and-limitsread 2026-11-08

Mike shows plan cards to approve or decline before work starts, Sarah's SEO pages wait for an approved plan, and Adrian shows two owner approval cards, a real cost warning and then the campaign plan, before creating a campaign that starts paused, with high risk Google recommendations needing a separate confirmation. Connectors may confirm before writing outside Atoms, edits arrive as a diff to submit or discard, and version history restores any earlier state. Budget changes on an existing campaign can skip a new plan card.

Sourcehelp.atoms.dev/en/articles/12129582-ads-agentread 2026-11-08

Accounts sign in by email and password, magic link or Google, and workspaces have Owner and Editor roles, with only the owner able to connect integrations such as Google Ads. Atoms says it does not use personal data or user content to train or improve AI models and hosts in the United States, but it names no SOC 2, ISO or other certification, no single sign on and no subprocessors, and describes encryption only in general terms.

Sourceatoms.dev/privacy-policyread 2026-11-08

Each project shows a terminal panel, console logs and an issue report card with stack traces, the credits each message used and cloud and AI usage refreshed daily. Jobs keep a run history, and the Marketing module brings GA4, Search Console and Google Ads results into dashboards. Atoms offers no audit log, export of agent actions or tracing beyond what the project chat itself shows.

Sourcehelp.atoms.dev/en/articles/12129554-app-viewerread 2026-11-08

Every agent reads the project state and the previous agent's output, long chats are organized automatically to keep context, and a memory compression experiment runs for some groups. API keys can be saved to a library and reused across projects, but no memory of a builder's preferences or past corrections carries from one project to the next or appears where a person can view or delete it.

Sourceatoms.dev/ai-agents/team-leader-agentread 2026-11-08

Atoms hosts published apps in the United States on its own domain, with development and production environments and custom domains on paid plans, and billing continues while an app is paused. A project can run on the customer's own Supabase organization instead of Atoms Cloud, which puts the backend in an account the customer controls, while ZIP or GitHub export is a way off the platform rather than another place to run. No region choice or self hosting exists.

Sourcehelp.atoms.dev/en/articles/12129577-publish-your-projectread 2026-11-08

Eight named agents come ready to work as a fixed team, from the team leader, product manager, architect and engineer to agents for deep research, data analysis, SEO and Google Ads, and custom agents cannot be added. Templates, themes, a Discover gallery and remixing of public projects give starting points, and output types cover web apps, Android apps, video, slides, documents and reports.

Sourceatoms.dev/ai-agentsread 2026-11-08

Agents start from a chat message, an at mention or a Marketing shortcut such as Optimize SEO, Create ads or Analyze, which sends a request into the chat. Atoms Cloud Jobs run a built app's own routes on a schedule with run history, but those call the app's code rather than an agent, and no schedule, event or API starts an agent, so the research, SEO and ads agents work only when someone asks.

Sourcehelp.atoms.dev/en/articles/12129619-jobsread 2026-11-08

A model selector offers Auto or a chosen model across Anthropic, OpenAI, Google, DeepSeek, Qwen and Zhipu, from Claude Opus 5 and GPT 5.6 Sol to Gemini 3.7 Flash and GLM 5.2, and only an admin can change a workspace's default model. Custom Race Mode runs up to four models on the same task in parallel. There is no bring your own key for the agents themselves, while built apps call AI with no keys and bill a separate wallet.

Sourcehelp.atoms.dev/en/articles/12129549-choose-an-llmread 2026-11-08

No public API, SDK, command line tool or MCP server is offered for the builder or its agents, and outside tools reach Atoms only through its own connectors and a Discord bot. A separate Atoms Router site presents an AI API gateway, but it shows no guides, pricing or link to building apps, so nothing lets another program create projects, start agents or read their results.

Sourcehelp.atoms.dev/en/articles/12129568-integrationsread 2026-11-08

An issue report card catches runtime errors with stack traces and a Resolve button sends them back to the engineer agent, terminal errors are detected automatically, and a security scan runs before each publish, though a builder can choose Publish Anyway. Atoms writes and runs no automated test suite that a person can see, so checks stop at error repair and the scan before an app goes live.

Sourcehelp.atoms.dev/en/articles/12129550-project-chatread 2026-11-08

A page validation agent opens the generated app in a browser, clicks through buttons, forms and navigation, takes screenshots and scores what works, and tool cards show the agent reviewing the rendered page as it builds. Changelog fixes for missed clicks, screenshot failures and accidental filling of login and sign up forms describe how it works on real pages, and it covers only the app being built, never outside sites or a desktop.

Sourcehelp.atoms.dev/en/articles/12129614-changelogread 2026-11-08