Your AI Agents Can Now Talk to 100+ APIs

The Problem We Solved

AI assistants are good at writing text and answering questions. But the moment you want one to actually do something, you hit a wall. "Send this email" turns into a custom integration project. "Create a Stripe invoice" means writing API glue code. Every new capability is a new engineering effort.

We wanted to fix that. We wanted agents that could reach out to real services and take real actions, without anyone writing a single line of integration code.

So we built the API Marketplace. It ships with 100+ OpenAPI specifications that get parsed at runtime. When you connect an integration to your agent, we read the spec, extract the available endpoints, generate AI agent-compatible tool definitions, and register them. Your agent can then call those tools during any conversation.

How It Works

Both the Assistant Builder and the Agent Builder follow a 5-step wizard. The API Marketplace lives at Step 4. The flow looks like this:

19 Categories, 100+ Integrations

We organized the integrations into 19 categories. Counts below.

Some popular integrations: SendGrid and Mailgun for email, Stripe for payments, Google Maps for location lookups, Mailchimp for marketing automation, GitHub for repository management, Jira for issue tracking, Notion for note-taking, NeuroGen Voice for SMS, Airtable for structured data, and Slack for messaging.

Security and Authentication

We support four authentication methods. The marketplace UI detects which one each integration needs and shows the right configuration panel.

There are also 4 integrations that need no authentication at all (public APIs like weather services).

Limits and Guardrails

We put limits in place to keep things sane and performant:

• MAX_INTEGRATIONS = 10: Each agent or assistant can connect up to 10 API integrations. This is enforced in both the Assistant Builder and Agent Builder.
• MAX_TOOLS = 20: Each integration can register up to 20 tools. APIs with more endpoints (like GitHub with 24 or Stripe with 30) get truncated to the 20 most commonly used endpoints.
• Tier gating: API integrations are available starting from the Starter tier ($47/month). The number of integrations you can use scales with your plan.
• Credit deduction: API calls made by your agent cost credits just like any other action. The credit cost depends on the LLM tokens used to process the API response.

A Real Example

Say you are building a customer support agent for your SaaS product. You want it to look up order status, send confirmation emails, and create support tickets.

In the API Marketplace, you would connect three integrations:

1. Stripe (finance category, bearer auth) for looking up payment and subscription data
2. SendGrid (communication category, bearer auth) for sending email confirmations and notifications
3. Jira (developer category, bearer auth) for creating and updating support tickets

You paste your API keys once. The agent gets tool definitions for things like search_charges, send_email, and create_issue. When a customer asks "Where is my order?", the agent calls the Stripe API, gets the payment status, and responds with real data. If the customer has a problem, the agent can create a Jira ticket and send a follow-up email, all in the same conversation.

No custom code. No webhook setup. No middleware.

What Is Coming Next

We are working on OAuth2 flows for QuickBooks and Shopify, which will bring our total to 77+ integrations. We are also building a visual tool for chaining API calls into multi-step workflows. Think "when a new Stripe charge comes in, look up the customer in HubSpot, send a personalized email via SendGrid, and log the event in Airtable." That is coming in March.

Back to Blog