You are in ChatGPT half the day anyway. Meanwhile your support inbox sits in another tab, and every time you want help with a ticket you copy the conversation out, paste it into ChatGPT, ask for a summary or a draft, then copy the answer back. It works, but it is manual, it loses context, and you quietly stop doing it by Thursday.
MCP (Model Context Protocol) removes the copy-paste step. ChatGPT can connect to remote MCP servers as custom connectors, and muro hosts one at a stable URL. Once they are wired together, ChatGPT can read your conversations, search your contacts, draft and send replies, and triage the whole inbox, all from the chat box you already live in.
What you can do once it is connected
- →Ask "what came in overnight?" and get an answer pulled from your actual conversations, not a hallucinated guess.
- →Summarize a long, messy thread before you wade into it yourself.
- →Draft a reply in your voice, edit it in the chat, then send it to the visitor without switching tabs.
- →Triage in bulk: close resolved threads, tag feature requests, assign the billing questions to your cofounder.
- →Pull a contact's history before a call: who they are, what they asked before, when they were last seen.
None of this needs webhooks or a line of glue code. The MCP server wraps the same operations as muro's REST API (the API docs cover the raw endpoints), so anything you could script, ChatGPT can now do conversationally.
What you need
- →A muro workspace with at least one site receiving conversations. The 14-day trial needs no card.
- →A paid ChatGPT plan that supports custom connectors. OpenAI has moved this feature between plans and settings screens more than once, so check your own plan rather than trusting anyone's screenshot, including this article's description.
- →A muro API key, which takes about a minute to create.
Step 1: create a scoped API key in muro
In the muro dashboard, open the API keys section of your workspace settings and create a new key. Two decisions matter here.
Scope: read, or read plus write. Every muro key carries scopes. read lets a client list and read conversations, contacts, and site config. write additionally lets it send replies, add internal notes, triage conversations, and change widget settings. Start with a read-only key. You lose nothing except the ability for ChatGPT to act without you, which on day one is a feature, not a limitation.
Workspace-wide or per-site. Keys can be scoped to a single site. A site-scoped key only ever sees that one site's conversations, contacts, and config, even through MCP. If you run several projects from one workspace and only want ChatGPT touching one of them, scope the key down to that site.
Copy the key somewhere safe (it looks like mr_live_…). muro shows it once.
Step 2: add the connector in ChatGPT
This is the part where I have to stay vague on purpose. OpenAI has renamed and moved the connector settings several times, so any precise click path would be stale within a month. The shape of the flow has stayed stable:
- 01In ChatGPT's settings, find the section for connectors (it has also lived under apps and beta features).
- 02Enable developer mode if your plan gates custom connectors behind it. Remote MCP servers with write tools usually are.
- 03Add a new custom connector and paste the server URL:
https://muro.chat/api/mcp. - 04For authentication, supply your muro API key as a Bearer token if the form accepts an API key or a custom header. The auth options ChatGPT offers vary by plan and rollout.
- 05Save, start a new chat, and enable the connector for that conversation.
If your version of the connector form only offers OAuth or no authentication, do not force it. Any MCP client can talk to the same server (more on that below), and if you are building on OpenAI's API rather than chatting in the app, their remote MCP tool config accepts custom headers, so you can pass the Bearer token directly there.
To confirm the connection works, ask something harmless: "List my five most recent conversations." ChatGPT should call list_conversations and come back with real threads from your inbox. If you get an authorization error instead, the key is wrong or picked up stray whitespace; muro answers with a clear 401 in that case, not a silent empty list.
What the muro MCP server exposes
The server speaks standard MCP (Streamable HTTP with JSON responses) and exposes eight tools. Five need the read scope:
- →
list_conversations: your conversations, newest activity first, filterable by status (open, closed, snoozed) and full-text search across contact names, emails, and message text. - →
get_conversation: one thread in full, with the contact, status, tags, and its recent messages in order. - →
list_contacts: the people who have talked to you, with name, email, and last-seen. - →
get_contact: one visitor and their metadata. - →
list_sites: your sites and each widget's current config: name, color, position, welcome message, agent name, language.
Three need the write scope:
- →
send_reply: send an agent reply the visitor sees in the chat widget, or add a private internal note that teammates see and the visitor does not. - →
update_conversation: set status, replace tags, assign a teammate, mark read. Closing a conversation resolves it. - →
update_site: change widget config, from the accent color and launcher position to the welcome message and whether visitors must give an email before chatting.
Because these wrap the same helpers as the REST API, everything an MCP tool does shows up in your inbox like any other agent action. A reply sent from ChatGPT is a normal agent reply, attributed to the API key that sent it, so you can always tell after the fact which messages a model wrote.
Prompts that actually earn their keep
- →"Summarize every open conversation and flag any that mention refunds or bugs."
- →"Read the latest thread from Anna and draft a reply explaining how CSV export works. Two short paragraphs, no exclamation marks."
- →"Which contacts keep coming back? I want to know who my heaviest users are before I plan the next release."
- →"Tag every open conversation that mentions pricing with pricing-feedback, then list them for me." (needs write)
- →"Close every conversation where the visitor's last message was just a thank-you." (needs write)
The pattern: anything that requires reading several threads and holding them in your head at once is where a connected model beats you on speed. You stay the editor. It does the reading.
The honest footnote: any MCP client works
Nothing here is ChatGPT-specific on muro's side. The server is plain MCP over HTTP with Bearer auth, so Claude Desktop, Claude Code, Cursor, Windsurf, or an agent you wrote yourself all connect to exactly the same URL with exactly the same key. For clients that speak stdio rather than remote HTTP, mcp-remote bridges it in one line:
bashnpx mcp-remote https://muro.chat/api/mcp --header "Authorization: Bearer mr_live_YOUR_KEY"Same URL, same key, same eight tools. If you switch models next quarter, your support integration comes with you. That is the quiet argument for MCP over building against any single vendor's plugin system: the work you do today is not a bet on one company's roadmap.
What this costs on the muro side
MCP access is included in both plans, not an add-on. Solo is $19/mo for up to 2 projects, Fleet is $59/mo for unlimited projects, and both come with unlimited agents since pricing is per project count, never per seat. One nuance worth knowing: connector calls run on ChatGPT's model, so they spend whatever plan you have with OpenAI, not your muro AI credits. Those credits stay reserved for muro's own AI, the widget-side auto-replies that answer visitors from your published help articles and hand off to a human when unsure.
If you want this wired up before lunch: create a workspace, add your site, paste the widget snippet, and import your docs URL so muro's AI handles the easy questions on its own. Then mint a read-only API key, add https://muro.chat/api/mcp as a ChatGPT connector, and ask it what came in today. The copy-paste era of your support inbox ends there.