Skip to content

Connect an AI agent ​

You can let Claude Code, Claude Desktop or Cursor read and act in Connectdary for you over MCP, in only the conversations you choose.

An agent acts as you, so its messages show as yours with "via Claude Code" beside the time. It never reads anything you didn't pick, and you can stop it any time. MCP needs the Team plan, and guests and people from another company can't connect agents.

The address of Connectdary's MCP server is:

text
https://connect.example.com/api/mcp

Two ways in ​

  • Sign in from the agent. You add the address to your agent, it opens Connectdary in your browser, and you pick what it may read. Nothing to copy. This is the easiest way.
  • An agent token. You make a token in Connectdary and paste it into the agent's config. Use this for tools that can't open a browser, or when you want a token that never expires.

Both give the agent the same kind of access: an agent token, tied to you, limited to the conversations you chose, and usable only over MCP. It doesn't work with the public API.

Sign in from the agent ​

  1. Add Connectdary to your agent (see the next sections).
  2. The agent opens Connectdary in your browser. Sign in if you aren't.
  3. Pick the conversations it may read. None are chosen at first.
  4. Read what it won't be able to do, and decide on Ask me before it posts (see below).
  5. Choose Allow.

Connectdary makes an agent token named after the client, for example "Claude Code". It is listed in Settings › Developer.

Make a token by hand ​

  1. Open Settings › Developer and choose New token.
  2. Choose For an AI agent, and give it a name.
  3. Pick the conversations it may read. An agent token needs at least one.
  4. Decide on Ask me before it posts, and create the token.
  5. Copy the secret. It is shown once. Keep it in an environment variable such as CONNECT_AGENT_TOKEN, not in a file you share.

Claude Code ​

Sign in from the agent:

bash
claude mcp add --transport http connect https://connect.example.com/api/mcp

Then run /mcp inside Claude Code, pick connect and choose to authenticate. Your browser opens Connectdary's consent screen.

With an agent token instead:

bash
claude mcp add --transport http connect https://connect.example.com/api/mcp \
  --header "Authorization: Bearer $CONNECT_AGENT_TOKEN"

Claude Desktop ​

Sign in from the agent: open Settings › Connectors, choose to add a custom connector, and enter https://connect.example.com/api/mcp. Claude opens Connectdary in your browser.

With an agent token, edit claude_desktop_config.json (Settings › Developer › Edit Config). Claude Desktop starts local programs, so the small mcp-remote helper carries the token to Connectdary:

json
{
  "mcpServers": {
    "connect": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://connect.example.com/api/mcp",
        "--header",
        "Authorization:${CONNECT_AUTH}"
      ],
      "env": { "CONNECT_AUTH": "Bearer cnct_paste_your_token_here" }
    }
  }
}

Restart Claude Desktop after saving.

Cursor ​

Add this to ~/.cursor/mcp.json (or .cursor/mcp.json in one project). With no headers, Cursor signs in through your browser:

json
{
  "mcpServers": {
    "connect": {
      "url": "https://connect.example.com/api/mcp",
      "headers": { "Authorization": "Bearer ${env:CONNECT_AGENT_TOKEN}" }
    }
  }
}

Choosing conversations ​

The agent can read only the conversations you picked, and a private channel only if you are in it. Ask it "who am I and what can you see?" and it uses whoami and list_conversations to tell you.

To change them later, open the token's row in Settings › Developer. A token can also be narrowed there with Edit scopes, and the change applies on the agent's next request, with no new secret.

Ask me before it posts ​

This is one switch on the consent screen and on the token. It is off at first.

  • Off: the agent acts at once as you, for every kind of action your admins have turned on. Each action shows in You › Assistant with the agent's name, and most can be undone there.
  • On: Claude Code asks you in its own prompt before it sends or schedules a message, edits or deletes one, reacts, starts a topic, starts a poll or resolves a topic. Tasks, reminders, sections and drafts still need no prompt. Other agents follow their own approval settings.

If the agent can read a direct message, it can't post where others see it, and it is told why. Picking a DM turns the switch off for that reason. Admins can switch off whole kinds of agent actions in Settings › AI.

What the agent gets ​

Connectdary gives the agent these tools. Which ones it sees depends on what the token allows. Anything people wrote comes back marked as data, so a message can't give the agent orders.

Read

  • whoami: who the agent acts as, the workspace and the conversations it may read.
  • list_conversations: channels and DMs it may see, a page at a time.
  • list_topics: topics in one conversation.
  • read_messages: the newest messages in a topic, up to 50, with paging back.
  • read_thread: a message and its newest replies.
  • read_inbox: your mentions, replies and unread items in the chosen conversations.
  • list_members: who is in a conversation (names, no email addresses).
  • list_reactions: who reacted to a message, and with what.
  • read_file: a file's details and its text, up to 20,000 characters.
  • search: messages matching words, up to 20.
  • list_tasks: your tasks, open by default, or done ones.

Write

  • message.send: send a message in a topic, or a reply in a thread.
  • message.edit and message.delete: change or remove a message the agent sent for you.
  • message.schedule: schedule a message for later.
  • reply.draft: put a reply in your composer for you to edit and send. Never sent.
  • topic.create, topic.rename, topic.resolve and topic.reopen: manage topics.
  • poll.start: start a poll in a topic.
  • checklist.set: tick or untick an item in a checklist the agent sent.
  • reaction.add and reaction.remove: react to a message, or take the reaction back.
  • task.create, task.complete and task.reopen: turn a message into a task, and finish or reopen it.
  • reminder.set: remind you of something at a time.
  • section.file: move a conversation into one of your own sections on Home.

Try it ​

Start with a read, and add a write once you trust it.

  • "Summarise what I missed in #launch since yesterday, and list anything waiting on me."
  • "Read the topic 'Pricing page copy' in #design and draft a reply to Ben that answers his last question. Put it in my composer, don't send it."
  • "Find where we decided on the new onboarding emails, and give me the link to that message."
  • "Look at my inbox and create a task for each thing someone asked me to do, with the message it came from."

Each prompt only reaches the conversations you picked. If it asks about a channel you didn't pick, the agent finds nothing there.

See what it read ​

Open You › Assistant to see what your agents did, and Settings › Developer to see what they read: reads per day and conversation for the last 30 days.

Stop an agent ​

Open Settings › Developer, find the agent's token and choose Revoke. The agent is stopped on its next request, and its open MCP session ends. To let it back in, connect it again. Deactivating your account stops your agents too.