API layer & connectors

Connector platforms sit between an AI agent and the SaaS APIs it needs: Gmail, GitHub, Slack, CRMs and so on. They give the agent tool definitions it can call, and they hold each end user's credentials so the agent can act for that user. Projects differ in four main ways. They define integrations either as declarative JSON or as hand-written servers. They expose tools to the agent through REST function calling, MCP or framework SDKs, and many handle large catalogs with tool search. Their credential storage ranges from a token file on disk to multi-tenant, KMS-encrypted linked accounts. And they differ in how much needs a vendor's cloud. When you choose one, check which OAuth flows you can run without the vendor and whether credentials are refreshed when a tool runs. Also check whether retries, rate limits and event triggers exist in the code or only in the hosted product.