WeWeb

WeWeb AI and outside agents working through MCP build apps against Supabase, Xano, Airtable, Make, Twilio, Slack, Google Sheets, Stripe, Resend, SendGrid, Segment, Cloudinary and other connected services, reading a Xano project's structure to build on it. Apps they build call those services and any REST or GraphQL API through HTTP request actions, connect to PostgreSQL, MySQL and other databases, and take Stripe, WhatsApp and Telegram events. Older Xano and Supabase plugin actions stay out of reach of MCP until a project moves to the newer integrations.

Sourcedocs.weweb.io/integrations/index.htmlread 2026-11-08

WeWeb AI runs in Build, Plan, Ask and focused edit modes for workflows, components and formulas, switching between them on its own, and Plan mode lays out the work and creates nothing until the user approves it. It asks clarifying questions, shows a run summary and live steps as it works, runs on a three agent architecture, and handles project tasks such as reading deployment history, explaining a failed publish, making and restoring backups and duplicating projects. Apps it builds get backend workflows with branches, loops, error handling and streaming responses.

Sourcedocs.weweb.io/weweb-ai/ai-modes.htmlread 2026-11-08

Every request carries the project's library, and a context manager adds pages, workflows, data, auth settings and environment variables, along with any selected elements and up to ten attached images, PDFs or text files of 3.8 MB each. Agents working through MCP pull WeWeb's documentation and integration skills with tools of their own. Context is assembled for each request, with no index or knowledge base kept over the customer's own documents.

Sourcedocs.weweb.io/weweb-ai/passing-context-to-ai.htmlread 2026-11-08

Plan mode waits for Approve and build or Ask for changes, the AI asks permission before acting with Yes, Yes to All, No or Stop, and any run can be stopped. Pro and Partner workspaces can lock resources, one by one or by category, so agents can only read them. Publishing is a step people take, from the editor to staging and production, with each publication versioned by author, message and logs and restorable, database structure changes need approval, custom roles separate staging and production rights, and editor backups roll a project back.

Sourcedocs.weweb.io/weweb-ai/protecting-resources-from-ai.htmlread 2026-11-08

Enterprise workspaces sign in through Google Workspace or Microsoft Entra ID with directory and group sync every few minutes, and custom roles and groups grant permissions per project, connection and admin area, including a permission to choose resources the AI must not change.

Two factor sign in comes on request under an Enterprise agreement, data is encrypted in transit and at rest, and the AI sees only development content, never staging or production data, under provider terms that bar training. WeWeb holds no SOC 2 or ISO 27001 certification of its own yet and is not built for protected health information.

Sourcedocs.weweb.io/settings-billing-code-export/workspace-settings/single-sign-on.htmlread 2026-11-08

A token gauge shows the AI allowance and workspace usage, with alerts at 75% and 90% and over the limit, and AI Gateway usage records the model, cost and triggering workflow for each call under a monthly cap. Backend request logs run per environment, workflow logs sit in the editor and reach agents through an MCP tool, and every publication records its author, message and logs. There is no audit log of who changed what in a workspace.

Sourcedocs.weweb.io/database-and-apis/logs.htmlread 2026-11-08

Chats keep a history the user can return to or start fresh from, the effort level is saved per project, and editor preferences are saved per browser. Design system guidelines that agents read and edit through MCP act as standing instructions for a project. WeWeb keeps no memory that the AI writes and recalls across chats, so what it learns in one chat reaches the next only when someone writes it into those guidelines.

Sourcedocs.weweb.io/weweb-ai/intro-to-weweb-ai.htmlread 2026-11-08

The backend runs on AWS with Neon as the managed database, in a US, EU or APAC region chosen when it is set up and fixed afterwards, served through an AWS CDN, while the editor itself runs in the US. From Essential up, the frontend and backend export as code with GitHub sync to host on any cloud, on premises or without internet access, with guides for Cloudflare and AWS, and an Enterprise API automates deploys and exports.

Sourcedocs.weweb.io/settings-billing-code-export/hosting-and-code-export.htmlread 2026-11-08

About 29 templates across eight categories start projects, including a ChatGPT style chat app, and a marketplace offers further templates and UI kits, some free and some paid. Integration templates cover Stripe checkout, file storage and a user API, and an AI Chat element, marked new, drops a chat interface into an app. WeWeb offers no prebuilt AI agents for customers to adopt.

Sourceweweb.io/templatesread 2026-11-08

Work reaches WeWeb AI from the editor chat, or through the MCP server from ChatGPT, Claude Desktop, Claude Code, Cursor, VS Code, Codex or Antigravity, always started by a person, and no API or schedule invokes the AI on its own. Apps it builds respond to UI events, sign up and sign in events, API endpoints, Stripe events and messaging webhooks, and run one time or recurring scheduled tasks on paid plans, which the AI and MCP can create, inspect and clean up but cannot pause or resume.

Sourcedocs.weweb.io/weweb-mcp/installing-weweb-mcp.htmlread 2026-11-08

WeWeb AI runs on models from Google Gemini through Vertex AI, Azure OpenAI, Anthropic and OpenAI, with WeWeb choosing among them, and customers set only an effort level for each request. There is no way to bring a key for the editor's AI, and teams that want more capacity can connect their own agent through MCP instead. In the apps people build, the AI Gateway runs Gemini by default and takes the customer's own OpenAI or Anthropic key per environment.

Sourcedpa.weweb.ioread 2026-11-08

The MCP server signs in through the browser and exposes about 159 tools that create and duplicate projects and create, change or delete pages, elements, components, design systems, variables, formulas, workflows, tables and rows, integrations, auth users and roles, scheduled tasks, environment variables, storage and backups. A single command adds it to Claude Code, and MCP calls draw no WeWeb AI tokens. Enterprise adds a public API for deploying and exporting projects.

Sourcedocs.weweb.io/weweb-mcp/available-tools.htmlread 2026-11-08

WeWeb AI tests the app it built by taking over the editor's preview: it opens pages, types and clicks, switches to mobile and tablet sizes, and reads errors, variables and workflow logs, checking for overflowing layouts, truncated text, poor contrast and workflows that fail silently. Test runs create real records prefixed qa_test_ that stay behind afterward, and testing works in every workspace but not through MCP.

Sourcedocs.weweb.io/weweb-ai/testing-your-app-with-ai.htmlread 2026-11-08

The AI's app testing runs inside the editor's preview of the app it is building, not in a separate browser, and no agent opens outside websites, fills in forms elsewhere or controls a desktop. Outside agents working through MCP change the project through its tools, not through the screen. Live previews and the editor's device switcher are there for people reviewing the app.

Sourcedocs.weweb.io/weweb-ai/testing-your-app-with-ai.htmlread 2026-11-08