Softr AI App Builder: Co-Builder, Agents, MCP and Credits

Softr's AI builds working business apps on your own data, but it does not hand you an app's code to take away. The AI Co-Builder assembles a database, pages, navigation, user groups with permissions and a theme from one prompt, using Softr's own blocks, and the Vibe Coding Block covers the parts those blocks miss. Every AI feature draws on one monthly credit pool per workspace: 5 credits on Free, 10 on Basic, 50 on Pro and 100 on Business. ^[1] Add-on packs start at +$10 per month for 100 credits on every plan, Free included. When the pool hits zero, AI pauses with no overage charge. The visual editor keeps working. Building through Claude or ChatGPT over Softr MCP spends no Softr credits at all.

Where the AI sits and what draws on credits

Softr spreads its AI across the studio where builders work and the finished app where users log in. Builders use the Co-Builder, the Vibe Coding Block and the MCP server. App users meet Ask AI and App agents. Data work runs through Database AI Agents and workflow AI steps. Everything except MCP, plain editing and own-key usage draws from the same pool, so a busy Ask AI button in a client portal can eat the credits a builder had set aside for new pages.

A first prompt with the Co-Builder

Since March 2026 the Co-Builder has been the front door. Describe the app, answer the two or three short questions it may ask, and it builds the schema first: tables, fields, relationships and sample data, then pages, navigation, user groups with permissions and a theme. ^[2] A detailed prompt can skip the questions. Nothing it creates is locked. The Co-Builder uses the same settings the visual editor exposes, so you can prompt one change, click the next, and resume an earlier AI conversation later.

Inside an existing app it reads the current page, the connected sources and the blocks already placed, then adds blocks, pages, charts, metrics or action buttons on request. That is where credits stretch furthest: Softr says a single block costs fewer credits than a full app, and the balance sits in workspace settings. Generate the skeleton once. Then switch to narrow requests. For anything no standard block covers, such as a Gantt chart, a pricing calculator or an org chart, the Co-Builder hands the job to a Vibe Coding component.

Know its limits. Conditional branching in multi-step forms still has to be set by hand in the Logic tab. The Co-Builder can create simple workflows, but complex multi-step automations may need manual adjustment, and system pages such as sign-in cannot receive AI-made custom blocks. A Plan mode appears on Softr's public roadmap as in progress, not shipped. Review the generated permissions page by page before real users sign in.

When standard blocks run out

The Vibe Coding Block, launched in December 2025, is the one place Softr's AI writes code. Each block is a TypeScript file exporting a React component that runs only in the browser, so server-side code and Node.js APIs are off the table. It uses Tailwind and shadcn/ui, can import public npm packages, and inherits the app's theme, data connections and permissions. It reads and writes Softr Databases and more than 15 external sources, and it can trigger a Softr Workflow through a webhook and wait for the reply.

Every re-prompt saves a version you can undo, restore or duplicate, and the code is editable by hand. Block code is checked before it is stored, so a broken attempt cannot replace the working version. The catch is permissions. Any code change resets the block's record actions to default visibility, so group restrictions must be set again after every recompile. The code also stays inside Softr: unlike code-generating builders such as Lovable, which can sync to Git, Softr exports neither the block nor the app, and offers no self-hosting.

AI that your app users actually touch

Ask AI, live since July 2025, is a chat button on dynamic blocks such as tables, lists, calendars and kanbans. It needs no API key, answers only from data the signed-in user may see, and lets the builder choose between the fields shown and the whole underlying table. Each query spends credits by response complexity and the data analyzed. Softr says it made Ask AI five times more cost-efficient in March 2026. It does not work on REST API sources or static blocks, and as of November 2026 Softr lists an app-wide version as coming soon.

Database AI Agents, out since September 2025, turn a Softr Databases field into one that categorizes, summarizes, extracts details from PDFs or pulls live web data. They fire on record creation, on specific field updates or on custom conditions, and each agent gets its own model from OpenAI, Anthropic, Gemini or Mistral. Use the conditions. Softr itself advises running agents with filters so credits are not burned on records that need no work. These agents cannot write to system fields or to formula and rollup fields.

Softr Agents, launched in August 2026, work across a whole app. App agents chat with users and act only as the signed-in user; Scheduled agents run unattended on a timetable. Both can search uploaded PDF, TXT or MD files, use native tools such as Gmail and Slack, and connect outside MCP servers, where write actions need approval by default. A Runs panel logs the model, tool calls and credits charged for every run. Start there when an agent's cost surprises you.

Workflow AI steps (write, summarize, categorize, or a custom prompt with JSON output) share the budget. On Softr's built-in models each step spends AI credits and also counts as one action against the monthly workflow limit, so a loop over hundreds of records presses on both meters. Bringing your own API key for agents and AI actions starts on Business; the provider then bills that usage. Action limits are covered in our Softr Workflows guide.

Building from Claude or ChatGPT

Softr MCP, launched in September 2026, lets Claude, Claude Code, ChatGPT, Codex or Cursor create databases, pages, Vibe Coding blocks and workflows, then publish the app. It is free on every plan, Free included, and spends no Softr AI credits, because the assistant runs on its own model and quota. Since October 2026 Softr has been an official connector in Claude and ChatGPT. Access is granted per area (Applications and Forms, Databases, Workflows) across the workspaces you choose, and the assistant acts as you.

That carries real risk. Through MCP an assistant can permanently delete records, fields, tables and whole databases, with no undo and no trash. ^[3] Only a View only connection can never delete. Publishing is one-way as well: the publish tool pushes every unpublished change, made by anyone, and no unpublish tool exists. Give an exploring assistant View only on Databases, take a manual snapshot first, and grant full access only for the session that needs it.

App as MCP is a separate server. Switched on per app, it lets your app users connect their own assistant to that one published app, runs every request as the signed-in user, and reaches only tables that the app's blocks display. The Softr CLI targets coding agents and terminals: it needs Node.js 22 or newer, covers the account, workflows and Softr Databases so far, and pushes workflow edits back as drafts.

Keeping the credit budget from stalling

Plan credits reset each month and never roll over; add-on credits roll over for one month. ^[4] Packs run from +$10 per month for 100 credits to +$500 per month for 5,000, at the same price on monthly and annual billing. By our arithmetic every pack works out to $10 a month per 100 credits, so a bigger pack buys no discount. Legacy plans cannot add packs. Because packs cost the same on every plan, move up a tier for a gate such as your own API key, not for its larger pool; plan prices sit in our Softr pricing guide.

Softr publishes no price list per AI action. Its only worked example is a Pricing-block AI edit at 0.89 credits, and the docs say usage depends on the model, the task type and the size of input and output. Measure before you buy. Build one feature, then read the workspace balance and the agents' Runs panel. When the pool empties, Co-Builder prompts, Ask AI, Database AI Agents and agents pause until credits renew or an add-on lands. Nothing is billed as overage.

The bottom line

Softr's AI is a quick route to a portal, CRM or internal tool on your own data, with users and permissions generated alongside the screens rather than bolted on later. It rewards discipline: one prompt for the skeleton, narrow prompts after, conditions on every database agent, and a pilot before buying packs. Its risks sit in recompiled Vibe Coding blocks and MCP connections with delete rights. Need code you can export or host yourself? Look elsewhere. Want AI help without owning a codebase? It fits.