Softr Agents
Build AI agents that chat with your users and automate tasks across your apps and data.
Overview
Softr Agents let you build an AI assistant that holds a conversation, follows your instructions, and takes real actions across your apps, databases, and connected services.
You don’t script every step. You describe what the agent is for and give it tools. The agent decides which tools to use, and in what order, to answer a question or finish a task. An agent can:
- Chat with your app users and respond in natural language.
- Search a knowledge base you upload (files or links) to give grounded answers.
- Take actions: read and write records in your data sources, send emails or Slack messages, search the web, and call external services.
- Remember relevant details about a user across the conversation.
- Run on a schedule, unattended, to handle recurring work.
Agent types
When you create an agent, you first choose how it runs. This can be changed later.
- App agent: Users chat with this agent from selected apps. Use it for support assistants, in-app helpers, and anything a person talks to directly.
- Scheduled agent: Runs on a recurring schedule, with no live user. Use it for recurring reports, data cleanup, sync jobs, and other background tasks.

The app as context for an App agent
When you connect an App agent to a Softr app, the app becomes a source of tools and data for the agent, the same way the App MCP Server exposes an app to external AI assistants. The agent reads and acts on your app’s data as the signed-in user who chats with it. It never acts as an admin or as the app itself.
So the agent can only do what that user can already do in the app:
- The blocks you build decide what’s possible. A list, grid, or table lets the agent read records. A form lets it create records (only the fields the form shows). A profile or inline editing lets it update. A delete action lets it delete.
- If no block shows a table, the agent can’t touch that table, even if the data source contains it.
- User groups, page visibility, and data restrictions still apply. Each user sees the same records and actions through the agent as they see in the app.
Step-by-step: Create your first agent
Create an agent
From your Softr dashboard, open Agents in the left sidebar. Click Create, or Create your first agent if you have none yet.

Choose an agent type
In the Choose an agent type dialog, pick App agent (users chat with it from your apps) or Scheduled agent (runs on a recurring schedule). Click Next. You can switch types later in the agent’s configuration.
Set up the agent profile
In the editor, give your agent a title and set up its profile:
- Model: Choose the AI model that powers the agent. Models are grouped by cost, so you can balance capability and price (see Choosing a model).
- Reasoning effort: For models that support it, choose how much the model thinks before it answers: Off, Low, Medium, or High. Higher effort gives better results on complex tasks, but it is slower and costs more.
- Instructions: Describe how the agent behaves and responds. State its goal, what it must always do, and what it must never do.

Connect it to your apps or add a schedule
Depending on the agent type:
- App agent: In the Connected apps card, click the + button to choose the Softr apps where the agent is available. You can Test the agent in each connected app.
- Scheduled agent: In the Schedules card, click the + button to create a recurring schedule. Set the frequency and a timezone. You can also add a message that starts each run.


Add tools
In the Tools card, click the + button to choose what the agent can do, such as read and write records in a data source, send an email, or post to Slack. Set each tool’s authentication. You can also require confirmation before the tool runs. See Tools & actions.

Add knowledge (optional)
In the Knowledge card, click the + button to upload files or insert links the agent can search when answering. See Knowledge.

Test and turn it on
Use the Test agent panel on the right to run the agent and watch its responses and tool calls. When you’re happy, enable the Agent enabled toggle and click Turn on to publish it. If you edit a published agent, a banner tells you that your draft differs from the live version. Click Update to publish your changes.

Choosing a model
Softr offers models from several providers, grouped by cost tier. Pick based on the task:
| When you need | Consider |
|---|---|
| The most demanding reasoning, coding, and agentic tool use | A top-tier model (for example a flagship GPT, Claude Opus, or Gemini Pro model) |
| A strong balance of capability, speed, and cost | A mid-tier model (for example Claude Sonnet or a Gemini Flash model) |
| Fast, cost-efficient handling of simple, high-volume tasks | A lightweight model (for example a mini, nano, or lite model, or Claude Haiku) |
Bring your own key (BYOK). On higher plans, you can use your own LLM provider API key. Your provider account then bills the agent usage, and it doesn’t use Softr AI credits.
Tools & actions
Tools are the capabilities an agent can use when it responds. Actions are tools tied to a specific integration, such as “Slack: send channel message” or “Google Sheets: create row.”
Native actions include:
- Softr: Read, create, update, and delete records.
- Google Sheets: Read, create, update, and delete rows.
- HubSpot: Read, create, update, and delete objects (for example contacts and deals).
- Gmail: Send emails and create drafts.
- Slack: Send channel messages, send direct messages, and find users.
- Send email: Send an email with Softr’s built-in email service.
When you add a tool, you configure:
- Authentication: Choose Per user authentication (uses each user’s own account or permissions) or Shared authentication (uses one shared account or connection).
- Ask for confirmation: When on, the agent shows a summary and waits for the user’s approval before it runs the action. Scheduled agents don’t have this option, because they run without a live user.
- Configure input fields: For each field, choose Agent decides (the agent fills it in at run time) or Custom (you set a fixed value the agent can’t change).
Connecting an MCP server
Agents can connect to external tool servers through the Model Context Protocol (MCP) to get more tools. Connect a predefined MCP server in a few clicks, or add a Custom MCP server with your own URL.
Predefined MCP servers
Softr offers ready-made connections with the server URL already set. You only authenticate. They include:
- Softr Databases: Read and manage your database tables.
- Airtable: Read and manage bases, tables, and records.
- Notion: Search and edit pages and databases.
- monday.com: Manage boards, items, and updates.
- Attio: Manage CRM records, lists, and notes.
- Apollo: Search contacts and enrich leads.
- Brevo: Manage contacts and email campaigns.
- Calendly and Cal.com: Schedule and manage bookings.
- Fathom and Granola: Access meeting recordings, summaries, and notes.
- Linear: Manage issues, projects, and cycles.
- Supabase: Query and manage your database.
- Clay: Enrich and research data.
- Apify: Run scrapers and automation actors.
- Replicate: Run and manage AI models.
- Atlassian: Work with Jira issues and Confluence pages.
Pick a predefined server, authorize it, and choose which of its tools the agent can use.
Custom MCP server
To connect your own MCP server, provide:
- Name: A label for the server, such as “Internal knowledge base”.
- Server URL: The MCP server endpoint.
- Authentication: OAuth, Headers (for example an
Authorizationheader), or none.
For each tool the server exposes, set a permission:
- Allow: Always allow the call, with no prompt.
- Ask: Require user approval before the call runs. Scheduled runs approve the call automatically.
- Deny: Never allow the call. The agent doesn’t get the tool.
Knowledge
Add files or links the agent can search when answering. Softr indexes uploaded files (PDF, TXT, MD) and links, so the agent can find and cite relevant passages.
Each source has a status while Softr processes it:
- Pending, then Indexing, then Ready: the agent can search the source.
- Failed: indexing didn’t finish. Remove the source and add it again.
Knowledge grounds an agent in your own content, such as product docs, policies, and FAQs. The agent answers from your material instead of guessing.
Testing & runs
Use the Test agent panel to run the agent and watch it work: its streamed response, the tools it calls, and the tasks it plans.
Open the Runs panel to review every run. Each run records:
- Model and Provider
- Duration and Tool calls
- Credits charged and Token usage (input, cached, output, total)
- Finish reason, plus a Run failed badge and a Show errors only filter for troubleshooting

Pricing and credits
Agent runs use AI credits. The credits a run uses depend on the model, the number and size of tool calls, knowledge search, and output length.
Every Softr plan, including Free, comes with a monthly amount of AI credits. If you use them all, AI features pause until your credits reset with your plan or you add more with a credit add-on. On higher plans, you can also bring your own API key, so your provider bills the usage. Learn more about AI credits
FAQ
What's the difference between Agents and Database AI Agents?
Agents (this page) are standalone AI assistants. They chat with your users or run on a schedule, and they take actions across your apps, data sources, and connected services.
Database AI Agents work inside a Softr Database table. They auto-fill, tag, summarize, or enrich individual fields when records are created or updated.
Both use AI credits, and you can use them together. For example, a Database AI Agent classifies a record, and a standalone Agent follows up based on that value.
Do I need to code to build an agent?
No. You set up agents visually: pick a model, write plain-language instructions, add tools and knowledge, and turn the agent on. A custom MCP server is optional. You only need it to give the agent your own external tools.
Can my app users chat with an agent?
Yes. Create an App agent and connect it to one or more Softr apps. Your users then chat with it inside the app. You choose the apps where it appears.
Can an agent run automatically without a user?
Yes. Create a Scheduled agent and add a recurring schedule with a timezone. It runs unattended on that schedule. Use it for reports, cleanup, and sync jobs.
Can I connect my own tools?
Yes. Add a Custom MCP server, or pick a predefined one, to give the agent external tools through the Model Context Protocol. You set an Allow, Ask, or Deny permission for each tool.
Which model should I pick?
Start with a balanced mid-tier model. Move up to a flagship model for demanding multi-step reasoning. Move down to a lightweight model for simple, high-volume tasks. See Choosing a model.
How are agents billed?
Agent runs use AI credits, which every plan includes. Usage depends on the model, tool calls, knowledge search, and output length. On higher plans you can bring your own API key. Learn more about AI credits