Every Aymar Agents integration — a native connector, an external MCP server you register yourself, or a copilot action — goes through the same four-step path before it reaches your AI agent or an identified caller. This page describes the mechanism; the Integration catalog describes each concrete connector.
1. Connection
Section titled “1. Connection”Nothing exists for your agent until you connect it. A first-party connector (HubSpot, say) connects with a token or credential you generate yourself in the external system — the Platform verifies that credential against the real system before storing anything; if verification fails, no credential is saved. An external MCP server (BYO-MCP) is registered from Settings → MCP servers, with the access credentials for the system you want to connect — the generic path for any system without a native connector yet.
2. Action-by-action allowlist, denied by default
Section titled “2. Action-by-action allowlist, denied by default”Connecting a system doesn’t turn anything on by itself. Every concrete action that system can offer — creating a contact, syncing a stage, adding a note — starts off; you turn each one on yourself, one at a time, from the integration’s own card. It’s the same principle that governs your agents’ resources (documents, templates, MCP servers): without explicit activation, it doesn’t exist.
3. Audience: Internal / Both
Section titled “3. Audience: Internal / Both”Every action declares whether it’s Internal (your team only, or an identified caller) or Both
(also an anonymous visitor on a public channel, such as your website’s widget). “Internal” is the default
for anything not explicitly declared. A Both-audience action that writes to the external system returns
an anonymous visitor only a minimal confirmation — never the full record, never a previous customer’s
data — minimized inside the action’s own executor, never left to the language model’s judgment.
4. The caller’s permissions
Section titled “4. The caller’s permissions”For an identified caller (a person with a session, an employee with their own channel, an API key), your account’s fine-grained permission catalog decides what they can do with that integration, intersected with the ceiling of their access medium — a key’s sealed scopes, or their channel’s cut-down. This layer is independent from audience: audience decides WHETHER an unidentified caller reaches the piece; permissions decide WHAT an identified one can do with it.
All four layers are always enforced on the server, never as a prompt instruction the language model could ignore.
MCP servers: the generic mechanism
Section titled “MCP servers: the generic mechanism”An MCP server is the bridge that connects your own business systems — your CRM, your catalog, your calendar — to your AI agent, so the AI can query them live while talking, instead of relying only on the documents uploaded to your knowledge base. There are two paths:
- A native connector (section 1) automatically registers a managed MCP server the moment you connect it — you don’t edit or delete it from Settings, it’s managed entirely from the integration itself.
- An external MCP server (BYO-MCP), which you register by hand in Settings → MCP servers, covers any system without a native connector yet, with more configuration steps in exchange for more reach.
Either way, you scope which agents can access the server with tags — denied by default, same as any other resource.
BYO-MCP: under your own responsibility
Section titled “BYO-MCP: under your own responsibility”Registering an external MCP server is an explicit option, clearly labeled as such: the system on the other side of that server is yours, contracted and managed by you, not your provider. Aymar Agents never offers an external MCP server “wide open” by default — every tool that server exposes still passes through layers 2-4 above before it reaches an agent or a caller.
Compliance: where your data ends up
Section titled “Compliance: where your data ends up”When an integration action sends data to an external system (a CRM, a spreadsheet, your own helpdesk), that destination system falls outside what Aymar Agents can guarantee or audit. You’re responsible for minimizing what you send, for your own information duties about that destination system, and for the fact that erasing an end customer on the Platform doesn’t automatically delete a copy already synced to your own system. See Data responsibility.
Pending walkthroughs
Section titled “Pending walkthroughs”Some integrations have their full connection procedure documented and verified with real screenshots; others don’t yet. Every catalog page says so honestly in its own status section — a step is never presented as verified when it isn’t.
See also
Section titled “See also”- Integration catalog — every connector, its actions and its audience.
- Settings and integrations — registering an MCP server of your own, step by step.
- Agent control: resources and permissions — the four layers in detail.
- Security and compliance — the same logic applied across the whole product.