Can Claude Cowork Send Text Messages? Connect It to Beam

Yes, once it is connected to a workspace that owns the actual sending infrastructure. Cowork cannot originate a text message on its own, the same way Claude Desktop or Claude.ai cannot either; none of Anthropic's clients ship with a phone line attached. What changes the answer from "draft only" to "actually sent" is a connector pointed at a platform like Beam, which supplies the business number, the carrier relationship, and the tools that do the sending and reading.

Once that connector exists, Cowork can call the same workspace tools any other connected client can: read a conversation's history, check which assigned numbers are reachable over iMessage versus falling back to SMS, draft a reply, and, with the right permission turned on, actually send it. That last part is worth repeating plainly, because it is the question underneath most of these searches: a correctly configured Cowork connection sends a message that lands on a real phone, not a preview Claude describes back to you in the chat.

1. In Beam, open Settings, then Developer access, then MCP & API. Create a connection, name it something recognizable like "Cowork", and choose only the permissions this particular workflow needs. Read and train are selected by default; sending, publishing, automation and booking each require an explicit toggle.
2. Click Create connection token and copy it immediately. It is shown once, and it cannot be displayed again after you leave that screen.
3. In Cowork, open Customize, then Connectors, then Add custom connector. Give it a name and enter the server URL, https://beam.aisync.link/mcp.
4. For authentication, choose No sign-in. Beam issues a Bearer token, not an OAuth flow, so the OAuth modes do not apply here.
5. Open the Request headers section and add a header named authorization. For the value, type the word Bearer, then a single space, then your token, exactly as shown; Anthropic's own documentation is specific that the scheme has to be included in the value you type, since Claude sends the header exactly as entered.
6. Mark that header Required, so the connection fails loudly instead of silently if the token value is ever missing.
7. Save the connector, then open a new chat and turn it on for that conversation from the + menu's Connectors list.
8. Before anything else, ask Cowork to call the workspace read tool and tell you the workspace name without changing anything, and confirm that name matches the client or business you meant to connect.

If step 5's Request headers section is nowhere to be found, stop there. Anthropic's documentation states this feature is in beta and available to a limited set of organizations, and no setting on your side turns it on early. Our bearer token connector guide covers that exact gap for Claude.ai in more depth, since Cowork's dialog follows the identical rollout.

Screenshot: the Cowork Add custom connector panel with the server URL field filled in, the No sign-in authentication option selected, and the Request headers section expanded showing an authorization header name with its value hidden behind dots.

What a connected Cowork session can actually touch depends entirely on the permissions you picked when you created the token in step 1, not on anything Cowork decides for itself.

• Read permission covers workspace_read, assistant_read, contacts_list, conversation_read, numbers_list, drafts_list, calendars_list and calendar_slots. This is what lets Cowork list recent conversations, pull up a contact's history, and check which of your assigned numbers are iMessage-enabled versus SMS-only through numbers_list, all without being able to change anything.
• Train permission adds draft_create and assistant_preview, for shaping the business's automated assistant without publishing changes live.
• Send permission adds message_send, the tool that puts a real text on a real phone. This is the one to add last, after you have verified the workspace and tested reads.
• Publish, automation and book permissions each add one further tool: draft_publish, conversation_pause, and appointment_book, respectively.

A token created with only read and train access cannot send anything no matter what Cowork is asked to do, since message_send simply is not on the list of tools it can call. That separation is worth treating as a real safety boundary, not a formality: verify a connection using read-only access first, and only add send once you trust what you are looking at.

Keep a short list and actually run through it, rather than trusting one clean-looking exchange:

Before calling a Cowork + Beam connection ready:
- Workspace read tool confirmed the intended workspace name, not a different one
- Request headers section was present and the header saved without an error
- A real test message sent to a contact you control, from this exact connector
- Message status checked inside Beam, not just a success reply inside Cowork
- Send permission added only after read-only checks passed
- Token scoped to this workflow, not shared with another client or job

The step people skip most often is the last line on Cowork's side: asking the agent to describe what it would send instead of actually sending it. A queued or accepted status in Beam is not the same thing as a message that reached a phone, so check the real delivery status before telling anyone on your team this connection is ready for a real customer.

If the workflow you are building also needs Cowork or another Claude client to place real phone calls, not just text, that is a separate connection on a separate platform. RizzDial documents the equivalent MCP setup for voice agents in its guide to adding phone calls to an AI agent over MCP, which follows a similar permission-first pattern before any call goes out.

For the complete tool catalog, every permission, and the full reasoning behind read-first testing, our iMessage MCP server guide goes deeper than the single connector covered here.