How to create an MCP connector (Company settings)

MCP connectors are added from Company settings → Integrations using New connector (wording in your tenant may match your theme). You step through a short wizard so Harriet can call an MCP server safely and attach it to skills.

AI-assisted setup (account owners)

If you are the account owner, you can also ask Harriet to research and create a connector in configuration mode:

• From Harriet chat, switch to configuration mode (owner-only) and describe the system you need, or paste links to official API/MCP documentation.
• During Endpoint AI / PLG onboarding, use Need a public MCP not in our catalog? to open the in-page Build MCP connector with Harriet chat (same configuration mode, without leaving onboarding). A link to the manual wizard remains for advanced setup.

Harriet will read your docs first, search the web only when needed, and refuse to propose a connector if it cannot find reliable source material. After you confirm a proposed configuration, Harriet creates the connector and can run a connectivity test (tool discovery). If the test fails, Harriet shows diagnostics and you can contact support@harriethq.com.

Credentials in chat: You may provide API keys, OAuth client secrets, or sandbox credential files in the conversation so the connector works immediately. Harriet does not read those secrets back in chat (only whether each credential was saved). You can rotate secrets later in Company settings → Integrations.

Wizard outline

1. Basics — Display name and how the connector should appear to admins.
2. Connection — Pick how Harriet reaches the server:
   • OpenAPI / Swagger — REST API via an OpenAPI spec URL and API base URL.
   • Native MCP server — HTTPS JSON-RPC MCP endpoint (MCP server URL required).
   • Sandboxed (uvx / npx) — Run a packaged MCP in an isolated environment (e.g. npx or uvx); you provide the sandbox command and arguments.
3. Authentication — Depends on type: no auth, basic auth, organization secrets, per-user OAuth (client ID when enabled), or patterns your deployment supports for sandboxed servers. Organization secrets are for IT only—Harriet stores them server-side and issues employees per-user, subscoped access (see MCP authorization and stored credentials).
4. Access — Which groups or document scopes may use this connector, consistent with your permissions model.

Save the connector, then use Sync tools (or equivalent) so Harriet discovers tools and you can set tool permissions (including confirmation gates for risky operations).

Endpoint AI review

If your organization uses Endpoint AI, new MCP connectors may start in a draft state and require submit for review before they are deployable. Reviewers use the Endpoint AI console review queue to approve, reject, or request changes. Link from the connector configuration UI when present.

After configuration

• Attach the connector to the appropriate skills (company settings or Endpoint AI skill authoring).
• Restrict which channels, workflows, or users may invoke those skills.

See also How do we connect external tools using MCP?, MCP authorization and stored credentials, and Using Skilify to build skills locally, submit for review, and provision them when authors submit package skills that reference connectors.

For OpenAPI / Swagger REST APIs and JavaScript request/response hooks, see How do we adapt an existing API into an MCP connector? and How can we restrict a shared-key MCP connector to each user's own data? For turning individual synced tools on or off by group, see How do we turn MCP tools on or off for specific groups?