Integrations

Integrations are how AgentSwarms talks to the rest of the world. The integrations page at /integrations is the central place to connect model providers, attach Model Context Protocol servers, enable pre-built SaaS connectors, configure webhooks, and set up email delivery. Every integration is scoped to your workspace, encrypted at rest with workspace- specific keys, and only ever decrypted inside the server-side runtime that actually uses it. None of the secrets you connect on this page are ever sent back to the browser.

Model providers

Model providers are the most commonly used integration on the platform, because every agent ultimately calls a model and every model lives behind a provider. AgentSwarms supports the long list of providers below out of the box, and connecting any of them is a matter of pasting an API key (or completing an OAuth flow where supported) and waiting a few seconds for the health check to pass.

• OpenAI via standard API keys, with full support for tool calling, structured output, and the embeddings endpoint. Organisation IDs and project IDs are configurable for accounts that use them.
• Anthropic via direct API keys with full support for tool calling. The Anthropic-specific cache- control headers are wired in automatically so prompt caching works without extra configuration.
• Google Gemini via direct API keys, with tool calling, structured output, and the multimodal endpoints (image and audio input) wired in.
• Grok (xAI) via direct API keys, including the long-context models when you have access to them.
• Azure OpenAI with per-resource endpoint, deployment name, and API version configuration. Useful when corporate policy requires Azure-hosted inference.
• AWS Bedrock with IAM-role configuration, regional endpoint selection, and access to every model family Bedrock exposes — Anthropic, Cohere, Meta, Mistral, Amazon Titan.
• Google Vertex AI with service-account authentication and project / location configuration, giving you access to Gemini and the third-party model garden on Vertex.
• OCI Generative AI for users in Oracle Cloud, with compartment and endpoint configuration.
• Alibaba Qwen via direct API keys, useful for accessing the Qwen model family at competitive pricing.
• Custom Ollama endpoints for anyone running open models on their own infrastructure. The integration accepts any Ollama-compatible URL and lists the models available at that endpoint automatically.
• vLLM endpoints for high-throughput self-hosted inference. The integration uses the OpenAI- compatible vLLM protocol so any model served by vLLM is immediately usable across the platform.

Every connected provider is health-checked a few times a day, with the status badge on the integrations page reflecting the result. A provider that goes unhealthy does not break anything: agents that were using it automatically fall back to their configured fallback model or to the AgentSwarms AI gateway, depending on how they were configured.

MCP servers

The Model Context Protocol page at /mcp lets you attach any MCP- compatible server to your workspace. Once attached, the platform probes the server, enumerates every tool the server exposes, and makes those tools available to attach to any agent in the Agent Builder. This is the primary route for connecting AgentSwarms to your internal systems — your CRM, your ticketing platform, your internal code search, your custom analytics warehouse — without having to write a bespoke HTTP tool for each one.

• Hosted MCP servers are attached by URL, with optional bearer-token authentication. The platform handles the protocol handshake and tool discovery for you.
• Self-hosted MCP servers running on your own infrastructure are attached the same way; we recommend running them behind authentication and using the bearer-token field rather than exposing them publicly.
• Per-tool enable / disable lets you choose to expose only a subset of a server's tools to your agents, which is useful when a server has a broad surface but you only want a small slice of it visible.

Standard connectors

Standard connectors are pre-built one-click integrations for common SaaS services. They differ from generic HTTP tools in that authentication is handled for you via OAuth where the service supports it, the tool schemas are pre-defined and kept up to date by us rather than maintained by you, and the connectors are tested against changes in the upstream APIs. Attaching a connector to an agent automatically exposes its tools to the agent without any further configuration.

The connector catalogue evolves continuously; the integrations page lists everything currently available. Whenever the same external service shows up in two or three unrelated user requests, we add a connector for it, with the result that the catalogue tracks real demand rather than our guesses about what would be useful.

Webhooks and public endpoints

Any swarm with a webhook trigger node automatically exposes a stable HTTPS endpoint that external systems can POST to. The endpoint URL is stable across renames of the project, which means you can configure third-party services against it without worrying that a future rename will break the integration. The same goes for any cron-driven external scheduler you might point at the platform.

• Stable URLs for both the published production deployment and the latest preview deployment are documented on the integrations page, so you can wire up external systems against either without ambiguity.
• HMAC signature verification is available for every public endpoint and can be enabled per endpoint. The shared secret is configured on the integrations page and never sent to the browser. Enabling signature verification is strongly recommended for any endpoint that performs a write or that is callable from the public internet.
• Rate limiting is applied to every public endpoint by default to protect both your workspace and the platform from a runaway external caller. The limit is per-endpoint and configurable upward when you have a legitimate need for higher throughput.

Email

Email delivery on AgentSwarms is handled by Lovable Email, which is wired in by default for all transactional and authentication messages: signup confirmations, password resets, magic links, budget alerts, contact-form replies, certification results, and any notifications from your own agents that choose to send email. The default sender domain is the platform's own; you can configure a custom domain from the account page if you want your agent-sent emails to come from your brand rather than ours.

• Custom sender domains are configured by adding DNS records (SPF, DKIM, DMARC) that we generate for you. The integrations page reports verification status in real time and tells you exactly which record is still missing if verification has not yet completed.
• Suppression management lets you view and manage the list of email addresses that have unsubscribed or hard-bounced, which is important both for deliverability and for compliance with anti-spam regulations.
• Template previews render every system- generated email (welcome, recovery, magic link, budget alert, certification result, contact confirmation) so you can see exactly what your users will receive before they receive it.