Set up connectors with your agent
Instead of stepping through the connector wizard, you can ask the agent in Harriet Desktop or in a cloud workspace to do it. Point it at the vendor's documentation and it reads the docs, proposes a connector, builds it, tests it and attaches it to a skill you can use.
Before you start
The agent does this work with the built-in Bot Configurator skill. Your admin adds it to your profile. In a cloud workspace, check that it is switched on under Workspace settings → Skills.
What the agent can do for you depends on your role:
| Your role | What happens when you ask for a new connector |
|---|---|
| Account owner or super admin | The agent creates the connector directly and can edit any connector. |
| Integration admin | The agent creates and edits connectors directly. |
| Anyone else, when your organization reviews new connectors | The agent submits the connector for review. It stays inactive until an admin approves it in the review queue. |
If your organization does not review new connectors and you hold none of these roles, the skill is not available to you. Ask an admin.
Point it at the documentation
You want your team's invoicing system available to your skills. Open a session in Harriet Desktop or your cloud workspace and give the agent what the vendor published:
- "Connect our CRM. The API docs are at this link. Start with read-only tools."
- "Here is the OpenAPI spec URL and the base URL for our internal ordering API. Make it a connector."
- "This is the README for the helpdesk's MCP package. Set it up for me."
Whatever you provide tells the agent which kind of connector to build:
| You give it | It builds |
|---|---|
| An OpenAPI or Swagger spec URL and the API base URL | A connector with one tool per operation |
| The address of a remote MCP server | A native MCP connector |
A package README with an npx or uvx command | A sandboxed connector that runs in Harriet's sandbox |
The agent reads your links first. It searches the public web only when what you gave it is not enough, or when you ask it to. If it cannot find usable official documentation, such as an install command, addresses and how sign-in works, it stops and tells you what to bring instead of guessing.
What the agent does
- Reads and plans. It works out the connection type, the address and the sign-in method. Where a server publishes its own sign-in details, it looks them up itself rather than asking you to paste them.
- Proposes, with sources. You see the configuration it intends to create and the pages it relied on. Nothing is created until you say yes.
- Creates and tests. It creates the connector and runs a test. If the test fails, it shows you the full diagnostic.
- Attaches it to a skill for you. It links the connector to a skill and assigns that skill to you, so the tools appear in your next session. It does not roll the skill out to anyone else. See How skills reach people for that step.
Credentials and sign-in
Never paste a token or client secret into the chat. When the connector needs one, the agent asks Harriet Desktop for a private form. You enter the value, select Save credentials, and it goes straight to encrypted storage. The agent only learns that it was saved. If the form is unavailable, use the connector's Integration settings instead.
For connectors where each person signs in to the vendor, the agent fills in the sign-in settings itself. It tells you the redirect address to register with the vendor and asks you only for the app's client ID and, through the private form, the client secret. Then you connect your own account:
- In Harriet Desktop the chat shows a Connect button.
- In a cloud workspace the agent gives you a sign-in link. Follow it and tell the agent when you have connected.
The agent never receives your token. The full steps for sandboxed connectors are in Hosted and sandboxed connectors.
Changing a connector that already exists
Ask for the change in plain words: "The ordering API moved to version 2. Update the connector." The agent reads the connector's current settings first and edits it. It does not create a second one.
Editing a live connector affects everyone who uses it immediately. The agent tells you exactly what will change and waits for your confirmation. To stage a change for review instead, ask it to propose an update. The live connector stays untouched until an admin approves the draft.
Get better results
- Give official documentation. A vendor's API reference or README beats a blog post.
- Say what you need to do. "Read invoices and create draft bills" lets the agent propose a narrower setup than "everything".
- Name the environment. Say whether the address is a sandbox or production account.
- Do one connector at a time. Test it with a harmless read before you build skills on top.
Limits to know
- Harriet reaches only public HTTPS addresses. A server on your office network needs a direct connection from Harriet Desktop.
- The agent configures existing servers and APIs. To have it write a new integration's code, use the Build integrations skill, described in Custom MCP servers.
- After it finishes, review which tools are switched on. See Tool permissions.
Once a connector works, your cloud workspace tasks can use it on a schedule or when a webhook arrives. See Orchestrating processes in cloud workspaces for examples.