How to connect an AI agent to your CRM

The agent never sees a password or a token. You connect the CRM to the platform once. The platform stores the credential encrypted, and when the agent calls a CRM action, the platform makes the API request on its behalf and hands back only the result.

In MyDeepAgents, HubSpot connects with a private app access token, and Salesforce with an OAuth sign-in (authorization code with PKCE), to a production org or a sandbox. Other CRMs include Pipedrive, Zoho CRM, Dynamics 365 Sales, Attio, Close, Copper, Freshsales and Zendesk Sell; the CRM category has more than 40 connectors in all. Browse them in integrations.

Credentials are opened only by the worker that makes the call, and each connector can reach only its own vendor’s hosts.

Each connector offers a fixed set of actions, and each action says whether it reads or writes. For the two most common CRMs:

• HubSpot

  Action

  Search contacts

  Reads or writes

  Reads

• HubSpot

  Action

  Create or update a contact

  Reads or writes

  Writes

• HubSpot

  Action

  Update a contact’s properties

  Reads or writes

  Writes

• HubSpot

  Action

  Create a deal

  Reads or writes

  Writes

• HubSpot

  Action

  Create a company

  Reads or writes

  Writes

• Salesforce

  Action

  Find records with SOQL

  Reads or writes

  Reads

• Salesforce

  Action

  Get a record

  Reads or writes

  Reads

• Salesforce

  Action

  Create a lead

  Reads or writes

  Writes

• Salesforce

  Action

  Create or update a contact

  Reads or writes

  Writes

• Salesforce

  Action

  Create an opportunity

  Reads or writes

  Writes

Neither connector has an action that deletes records, so an agent using them cannot delete anything in your CRM.

In MyDeepAgents:

1. 1Add a connection for your CRM in the console. For HubSpot, paste a private app access token; for Salesforce, choose production or sandbox and sign in.
2. 2Open the agent and add only the CRM actions it needs as tools.
3. 3Keep “ask first” on for actions that write. The console sets it for every new tool that writes, and you can switch it off for writes you are happy to let run on their own.
4. 4Write the instructions: search before creating, which fields to fill in, and what the agent must never change.
5. 5Try it in the test chat, against a sandbox where you can, and read each tool call’s input and output.
6. 6Publish it, and start it from where the work begins: a website chat, a workflow, or another agent.

• On the CRM’s side, give the token or the signed-in user only the access the agent needs.
• On the agent’s side, add only the actions it needs. An agent that only answers questions needs only the read actions.
• Fixed values: a tool can carry preset arguments, such as a pipeline or an owner, that always win over the model’s choice and are hidden from it.
• Fewer fields back: a tool can return only the fields you name, so the model never sees the rest of the record.
• Approvals on writes, so a person checks each change before it happens. See human-in-the-loop approvals.
• Data class: CRM actions are marked as handling personal data, so the agent’s compliance profile and policies apply to them.

Not directly yet. MyDeepAgents does not copy or sync your CRM data: every action reads or writes the CRM at the moment it runs, so the agent always works on current records. Workflows cannot yet start from CRM events by themselves.

Two patterns work today. A workflow on a schedule can run an agent that searches for new or changed records and deals with what it finds. Or any system that can call a web address can start a workflow through its webhook, signed with HMAC or carrying a token. See AI agents vs workflows.