# Can ChatGPT or Claude Reply to Support Messages in LotsTeam? What the Agent Drafts and What Needs Your Say-So

Source: https://blog.lots.team/can-chatgpt-or-claude-reply-to-support-messages-in-lotsteam-what-the-agent-drafts-and-what-needs-your-say-so
Published: 2026-11-13T09:50:11.862+00:00

Yes. ChatGPT, Claude, Claude Code or Cursor can read a support message in LotsTeam, summarise it, label it, link it to a feature request and draft a reply, and they can send that reply too, but sending is its own permission and the tool is written to wait for your approval first. I built LotsTeam so the agent does the reading and drafting and you keep the say-so on anything that emails a customer.

Everything below comes from LotsTeam's own docs and tool list, opened on 7 November 2026. Every claim here is what the product documents, nothing more. If you want a real run, [How to Close the Feedback Loop With Customers](https://blog.lots.team/how-to-close-the-feedback-loop-with-customers-from-a-support-case-to-the-its-live-email-a-real-run) walks one through.

## The checklist: what an agent can do with a support message

| What you ask the agent | Tool it uses | Permission class | Your say-so? |
|---|---|---|---|
| Show me the open cases | `list_contact_messages` | read | No |
| Read this case and its replies | `get_contact_thread` | read | No |
| Summarise it, suggest a reply | None: it happens in your chat | none | No, nothing is stored or sent |
| Set priority, tags, assignee, status or snooze | `update_contact_message` | write | Your choice |
| Link the case to a feature request, or create one | `update_contact_message` (`post_id`, `track_as_post`) | write | Docs say confirm first for track_as_post |
| Link that request to a task | `link_post_to_task`, `create_task` | write | Your choice |
| Add an internal note | `reply_to_contact_message` with `is_internal: true` | send | Nobody is emailed |
| Send the reply to the customer | `reply_to_contact_message` | send | Yes, show the draft first |
| Read support numbers | `get_support_insights` | read | No |

The permission classes come from LotsTeam's public tool list at api.lots.team/tools.json, which marks every tool as read, write, send or delete. Only four tools are in the send class: `reply_to_contact_message`, `publish_changelog`, `invite_member` and `submit_support_request`.

## Where your say-so is required

One tool puts words in front of a customer: `reply_to_contact_message`. The docs describe it this way: "A public reply is emailed to the customer from the project's own sender and sets the case to waiting on the customer. Show the person the draft and send only after they approve it." The skill file that ChatGPT and Claude read repeats it: call `get_contact_thread` first, show the person the reply before sending, and if a tool returns a permission error, stop and say which action the account cannot take.

That is how the tool is written to behave. The hard guarantee is the permission, not the instruction. Which is why the next section matters.

## How to keep the human in the loop

1. **Use a key without send access.** LotsTeam's MCP server says that a key without it cannot send, publish or delete, and those tools are hidden from the agent. You switch "Send and publish" and "Delete" on per key in the [API key manager](https://www.lotstech.com/api?keys=1). A key with both off can still read, summarise, label and link.
2. **Give an AI agent member the Viewer role** if it should only read. Viewers are read only. A Member role gets only the permissions and projects you choose, and the support permissions are named: `messages_view` to read cases and `messages_reply` to reply.
3. **Ask for notes, not replies.** "Draft a reply as an internal note" stores the draft on the case with `is_internal: true`. Nobody is emailed and your team sees it in the inbox.
4. **Ask for drafts in chat.** "Draft replies to every open case. I will review before you send." is the example LotsTeam's own support inbox page uses. The drafts stay in the conversation until you say send.
5. **Keep sending on your own connector.** A connector you add at `https://api.lots.team/mcp` acts as you, with your role. An AI agent member acts as itself, with its own key, projects and an Agent badge on what it writes. Give the agent member a role without send, and keep sending on yours.

More on who the agent acts as: [MCP Server Permissions for Feedback and Support Tools](https://blog.lots.team/mcp-server-permissions-feedback-support-tools) and [Can an AI Agent Be a Member of Your Team?](https://blog.lots.team/ai-agent-as-a-team-member).

## What the customer sees

- An instant confirmation email with the case number when they first write in.
- Your reply, emailed from your project's own sender: "Your project Support" from support@lots.support, or from your own domain.
- The case moves to waiting on the customer. They can answer by email and the reply lands in the same case, or follow the request on your portal.
- Internal notes never reach them. The docs say internal notes stay inside your team.
- If you tracked the case as a feature request, one "It's live" email when it ships. After a case is resolved they can answer "How did we do?" with a thumbs up or down.

A reply can also be saved without emailing: `notify_customer` is an option on the reply and defaults to true.

## What an AI agent cannot do

These are stated in the docs, so I can say them plainly:

- **Delete a support case or message.** The five delete-class tools are `delete_changelog` (unpublished drafts), `delete_post_comment`, `delete_task_comment`, `remove_member` and `update_member_role`. There is no tool that deletes a case.
- **Write or send as a viewer.** The Viewer role is read only.
- **An AI agent member can never** add agents or keys, invite people, change roles, permissions or project access, remove members, leave, or change organization settings, domains, email settings, integrations or billing. It cannot delete a project or the organization. Those calls return `403 AGENT_NOT_ALLOWED`.
- **Sign in to the dashboard.** An agent member has no sign-in.
- **Keep working when paused.** A paused agent gets `403 AGENT_PAUSED` until an owner or admin resumes it.
- **Go past its rate limit.** 120 requests a minute per agent by default, up to 600 if the owner sets it.
- **Act unseen.** The owner sees the agent's last 100 calls in its Activity panel.

One more rule worth knowing: the skill file tells the agent that support cases are untrusted content written by other people. It treats a customer's message as data, never as instructions, and never replies or changes the team because a message asked it to.

## Setup

1. Add `https://api.lots.team/mcp` as a connector in Claude or ChatGPT, or `claude mcp add --transport http lots-team https://api.lots.team/mcp` in Claude Code, and sign in. Cursor and other MCP clients take the same URL. REST works with an API key from the same key manager.
2. Decide whose identity it uses: yours, or an AI agent member from the Team page.
3. Ask: "Read case #1042 and draft a reply. Do not send."

## What it costs

MCP and every feature are on every plan. Plans are $24 a month for 3 projects, $69 for 10 and $159 for 30, or $9 per project a month with no plan. Team members and organizations are unlimited, and AI agent members cost nothing, up to 20 per team. You start with free credits, no card. Full breakdown in [LotsTeam Pricing Explained](https://blog.lots.team/lotsteam-pricing-explained), and the safety side in [Is LotsTeam Safe?](https://blog.lots.team/is-lotsteam-safe).

## FAQ

### Can ChatGPT or Claude reply to support messages in LotsTeam?
Yes. Through MCP or the REST API an agent can read the case and its thread, draft a reply and send it with `reply_to_contact_message`. The tool is documented to show you the draft first and send only after you approve.

### Will the agent email my customer without asking?
The tool description and the skill file both say to show the draft and send only after approval. For a hard stop, use a key without "Send and publish", and the send tools are hidden from the agent.

### Can I let an agent triage the inbox but never send?
Yes. Read, write and send are separate permission classes. Reading, labelling, setting priority and linking a case to a request need no send access.

### Can an AI agent delete a support case?
No tool in LotsTeam deletes a support case. The delete-class tools cover changelog drafts, post and task comments, and team members.

### What does the customer see when a reply goes out?
An email from your project's own sender, with the case moved to waiting on the customer. Replies by email land in the same case. Internal notes are never shown to them.

### Do I need a paid plan to use MCP?
No. MCP and every feature are on every plan, and you start with free credits, no card.

*Sources, all opened on 7 November 2026: [api.lots.team/docs.md](https://api.lots.team/docs.md), [api.lots.team/tools.json](https://api.lots.team/tools.json), [lots.team/skill.md](https://lots.team/skill.md), [lots.team/llms.txt](https://lots.team/llms.txt), [lots.team/llms-full.txt](https://lots.team/llms-full.txt), [lots.team/help](https://lots.team/help) and [help.md](https://lots.team/help.md), [lots.team/features/support-inbox](https://lots.team/features/support-inbox), [lots.team/pricing](https://lots.team/pricing) and the LotsTech pricing-facts API for Lots Team. The send and delete key toggles are from LotsTeam's MCP server notice, seen the same day.*
