Agent Handler runs three ways - same product surface, different end user and auth model. Most teams pick one; some run two in parallel.
Your product has an agent, and the people using it connect their own Salesforce, Slack, GitHub. The end user can be a customer’s user (B2B SaaS) or one of your own employees on an internal product. Either way, you never touch their credentials directly - Agent Handler holds them, scoped to one Registered User per person.
Pick this when:
Read Building an agent.
Your employees plug an MCP URL into their own AI surface - Claude, ChatGPT, Cursor, Claude Code, any MCP-compatible client - or into agents they build for their own work. Auth runs through your IdP (Okta, Azure AD, Google Workspace, whichever). SCIM provisions users into Agent Handler. IdP Groups map to Tool Packs.
Pick this when:
Read Context layer for employees.
Self-serve setup for an individual developer. You’re not shipping a product to customers, and you’re not standing up employee access for an org - you just want tools in Cursor or Claude Desktop on your own machine. Use the Merge CLI or the simplified MCP URL with your own login.
Pick this when:
Read Local development.
The two-question version: who is the end user, and who controls their identity?
If you’re not sure, start with Local development to feel out the surface, then move to Building an agent or the Context layer for employees once you know which shape your product is.