FlutterFlow

The agent works in outside accounts on the builder's behalf. Agent tools connect personal accounts in Airtable, ClickUp, Granola, Jotform, Linear, Miro, monday.com, Notion, Stripe, Tally and Wix, each action set to always run, ask first or never. Apps it builds connect to Stripe Checkout, AdMob, Resend, Airtable, Notion, Discord, Telegram and OpenAI, Anthropic or Gemini on the customer's key, and on request to Firebase, Supabase, Auth0 or a custom API. Store builds go to App Store Connect and Google Play through the customer's accounts.

Sourcedocs.flutterflow.io/ffx/agent-toolsread 2026-11-08

A first build runs as a plan: the agent proposes the app on a card, and once approved it plans, builds and checks each screen in turn, sets up backend tables, sign in and storage, adds starter data and an icon, then reviews the code it changed and fixes errors. Several threads can run at once, each on its own copy of the app and its data, and a push merges a thread back after checking that everything works and asking the person to settle any conflicts.

Sourcedocs.flutterflow.io/ffx/the-agentread 2026-11-08

Context is the thread itself, reread in full on every reply, plus up to six attached files of 4 MB each: images, text documents and fonts, with source code files refused. The agent also works from the app's own screens, backend tables and data. Studio's agent panel reads the project through a local workspace and accepts images. Neither agent keeps a searchable index of company documents or a knowledge base that persists outside the project.

Sourcedocs.flutterflow.io/ffx/attachmentsread 2026-11-08

The agent waits for approval before a first build, asking through a proposal card with Build it and Not now. Every thread works on its own copy of the app and its data, the agent never pushes on its own, and only the person who started the thread can push, after checks pass and conflicts are resolved. Actions in connected accounts can be set to ask first, and an approval expires after an hour. In Studio, changes from the agent panel are reviewed before they apply.

Sourcedocs.flutterflow.io/ffx/pushread 2026-11-08

Projects have Owner, Developer, Viewer and Billing roles, and only the person who started a thread can push it. Enterprise adds single sign on and audit logs. FlutterFlow holds SOC 2 Type II, audited each year, with a trust center run on Vanta, and a DPA, standard contractual clauses and representatives in the EU and UK cover GDPR. Its privacy policy says prompts and outputs are not used to train AI models and that its AI providers do not train on customer data either.

Sourcedocs.flutterflow.io/ffx/rolesread 2026-11-08

Each agent turn ends with a summary of what changed, and the screens tab marks the screens it touched. A backend activity view shows who pushed which change to tables and storage, day by day, though changes to individual records are not logged. Studio keeps snapshots, from an hour on Free to seven days on Max, records each agent push as a commit and offers an activity log through its command line tool, and Enterprise adds audit logs.

Sourcedocs.flutterflow.io/ffx/backend-activityread 2026-11-08

Each thread keeps its full conversation, which the agent rereads before every reply, so instructions given earlier in a thread carry through that piece of work. A new thread starts fresh from the pushed app. There is no project memory, saved rules file or store of preferences that the agent writes and recalls across threads, and Studio's agent panel works the same way in each desktop session.

Sourcedocs.flutterflow.io/ffx/threadsread 2026-11-08

FlutterFlow runs as a hosted service, processing data in the United States on Google Cloud, with no region choice for the agent or the backends it creates. Enterprise lists private cloud options without further detail. Apps it builds ship to the App Store and Google Play under the customer's own developer accounts, and Studio projects can export Flutter code to host elsewhere, while web agent apps live at an address FlutterFlow assigns.

Sourceflutterflow.io/pricingread 2026-11-08

The FlutterFlow marketplace offers whole app templates, individual widgets and custom plugins that a Studio project can start from or add to, so a team adopts prebuilt pieces instead of building every screen. The web agent opens with three ready starting prompts, for a finance app, a habit tracker and a marketplace, and Studio's agent panel can then work on any project a template creates.

Sourcemarketplace.flutterflow.io/homeread 2026-11-08

Agent work starts from a person: a prompt in a new thread, a message in an existing one, or a choice on a question card. When Apple sends a change request on a store submission, a Fix with the agent button opens a thread already filled in with the request. Studio's agent panel runs in the desktop app, and outside coding agents reach Studio projects through its MCP server and command line tool. Neither agent runs on a schedule, a webhook or an incoming event.

Sourcedocs.flutterflow.io/ffx/how-releasing-worksread 2026-11-08

The web agent runs on models FlutterFlow chooses and does not name, from AI providers that include OpenAI, Anthropic and Google, with Runware for images. Studio's agent panel works differently: it runs on the customer's own Claude, Codex, Antigravity or Gemini account, and the model and reasoning level can be picked for each message. That choice lives in the beta desktop panel, not in the web agent that FlutterFlow now leads with.

Sourceflutterflow.io/studio/airead 2026-11-08

Studio exposes an MCP server and a command line tool, set up with flutterflow ai init, that let Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode, Kiro and other agents read and edit FlutterFlow projects on paid plans, with plugins for Claude Code and Codex. A project API, labeled beta, reads and writes project files as YAML over REST. Deploying stays out of reach of outside agents, and the web agent has no API or MCP of its own.

Sourcedocs.flutterflow.io/flutterflow-cli/buildread 2026-11-08

Test Pilot, in Studio, writes tests in plain English and has an AI click through a web build of the app, reporting which tests pass; each project gets five free credits, then passes start at $5 a month for 100 credits. The web agent checks the code it changed for errors after each turn and runs checks before every push, and it says plainly when it cannot try a feature itself, such as maps.

Sourcedocs.flutterflow.io/testing/test-pilotread 2026-11-08

After changing how a screen looks, the web agent takes a screenshot of it and checks the result against the request, viewing its own preview without clicking through it. Studio's MCP server can capture screenshots of the builder canvas for outside agents. Clicking through a running app happens only inside Test Pilot runs, which a person starts from the testing tools in Studio.

Sourcedocs.flutterflow.io/ffx/the-agentread 2026-11-08