It is 8:45 on a Tuesday. Before you touch your code, you have to check what support brought in overnight: another tab, another dashboard, ten minutes of scrolling before you know whether anything is on fire. You do this every morning, for every project you run.
If you already live in Claude Desktop, you can skip the tab. muro ships a hosted MCP server, which means Claude Desktop can read your support conversations, draft replies, and close threads directly from the chat window. Setup is one API key and one JSON block in claude_desktop_config.json. This guide shows the exact config, then the three workflows that make it worth the ten minutes.
What the MCP connection actually does
MCP (Model Context Protocol) is the open standard from Anthropic that lets Claude call external tools. Claude Desktop supports it natively, and muro exposes your inbox as a set of MCP tools at https://muro.chat/api/mcp. The mcp-remote bridge connects the two, authenticated with your API key.
Once connected, Claude Desktop can work your inbox the same way you would in the dashboard:
- →Read conversations. List open threads, pull the full history of any conversation, and see who has been waiting longest.
- →Draft and send replies. Write a response into a thread, from your instructions and in your voice.
- →Update thread state. Close conversations that are resolved, so the open count tells the truth.
- →Answer questions about your queue. Volume by topic, repeated complaints, that one thread you half remember from last week.
Get a scoped API key first
In the muro dashboard, open your site settings and find the API keys section. Create a new key, and note two things before you paste it anywhere.
- →Keys are per site. A muro API key is scoped to a single site. The key in your Claude Desktop config can only touch that project's conversations, nothing else in your workspace. Three products means three keys.
- →Read and write are separate scopes. A read scope lets Claude list and summarize conversations but change nothing. A write scope also lets it post replies and close threads. Start read only. Run triage with it for a week, then swap in a write key once you trust the loop.
Configure claude_desktop_config.json
Claude Desktop loads MCP servers from a JSON config file. You can open it from inside the app (in the settings, look for the developer section and its edit config button) or edit the file directly:
- →macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - →Windows:
%APPDATA%\Claude\claude_desktop_config.json
Add an mcpServers entry that runs mcp-remote against the muro endpoint, passing your key as a Bearer header:
json{
"mcpServers": {
"muro": {
"command": "npx",
"args": [
"mcp-remote",
"https://muro.chat/api/mcp",
"--header",
"Authorization: Bearer mr_live_YOUR_KEY"
]
}
}
}The only prerequisite is Node.js 18 or newer, because npx downloads and runs the mcp-remote bridge the first time Claude Desktop starts it. No project setup, no install step, just Node on your PATH.
json{
"mcpServers": {
"muro": {
"command": "npx",
"args": [
"mcp-remote",
"https://muro.chat/api/mcp",
"--header",
"Authorization:${AUTH_HEADER}"
],
"env": {
"AUTH_HEADER": "Bearer mr_live_YOUR_KEY"
}
}
}
}Save, then restart Claude Desktop completely. On Windows, quit it from the system tray; closing the window is not enough. When it comes back, check the tools indicator near the chat input to confirm the muro server is listed, then verify with something simple: "List my open muro conversations." If Claude answers with real threads, you are wired up.
Workflow 1: morning triage
This is the workflow that pays for the setup on day one. Instead of opening the dashboard and scrolling, start the day with one prompt:
“Summarize what came in on muro overnight. Group by topic, flag anything angry or mentioning refunds or billing, and tell me which threads need me first.”
Claude lists the open conversations, reads the new messages, and hands back a ranked digest: two bug reports about the same webhook timeout, one pricing question, one happy customer, one thread that smells like churn. Follow up in the same window: "show me the full webhook thread" or "draft an answer for the pricing question."
The value is not that reading ten messages is hard. It is that the digest turns an anxious inbox check into a two minute status report, and you decide what deserves attention before anything grabs it.
Workflow 2: draft replies in your voice
Claude's default support voice is polite and bland. Fix that once per session: paste two or three of your best past replies and say "match this: short sentences, no corporate apology padding, sign off with my first name." Then work the queue:
“Draft a reply to the webhook timeout thread. We shipped a fix in 1.4.2 this morning. Ask them to retry, and offer to look at their delivery logs if it still fails.”
With a read only key, you copy the draft into the dashboard yourself. With a write key, you review it in Claude Desktop and tell it to send. Either way, keep the human review step. That is the same rule muro's built in AI follows: automation drafts, a person owns anything sensitive.
When a draft turns out to be an answer you will send weekly, store it as a saved reply in muro. The next occurrence becomes one click instead of one prompt.
Workflow 3: close resolved threads
Every inbox accumulates threads that are finished but still open. The customer said thanks, you moved on, and now the open counter lies to you every time you look at it. With a write scope, cleanup becomes a thirty second sweep:
“Go through my open muro conversations. List the ones where the issue looks resolved or the customer's last message was a thank you. Do not close anything until I confirm.”
The confirmation step is the whole trick. Claude proposes, you glance at the list, then approve the batch. Ten seconds of review prevents the one bad close on a thread that was actually still live.
Running more than one product
Because keys are scoped per site, multiple projects means multiple entries. Claude treats each one as its own tool set, so you can triage a single product or ask across all of them:
json{
"mcpServers": {
"muro-app": {
"command": "npx",
"args": ["mcp-remote", "https://muro.chat/api/mcp", "--header", "Authorization: Bearer mr_live_KEY_FOR_APP"]
},
"muro-devtool": {
"command": "npx",
"args": ["mcp-remote", "https://muro.chat/api/mcp", "--header", "Authorization: Bearer mr_live_KEY_FOR_DEVTOOL"]
}
}
}On the muro side this stays cheap: Solo is $19/mo for up to 2 projects, Fleet is $59/mo for unlimited projects, and both come with unlimited agents, no per seat pricing. Details on the pricing page.
How this fits with muro's built in AI
The MCP connection is for you, the operator. It is a different layer from muro's AI auto reply, which answers customers directly in the widget, grounded on your published help articles and product context, and hands off to a human when it is unsure (refunds, account issues, anger). Run both: the auto reply absorbs the repetitive questions in real time, and your Claude Desktop session works whatever queue remains, on your schedule.
The cost split is clean too. Triage and drafting in Claude Desktop run on your existing Claude plan, so they consume nothing on the muro side. muro's own automated replies draw from the managed credit pool (1,000 credits per month on Solo, 3,000 on Fleet, 1 credit = 1 AI action), or you can paste your own Anthropic API key and run unlimited AI at zero markup.
If Claude Desktop is already open on your machine, you are most of the way there. Create a muro site (14 day free trial, no card), paste the widget snippet on your page, and point the AI at your docs by importing your help center URL. Then add the config block above, restart Claude Desktop, and tomorrow at 8:45 ask it what came in overnight. One less tab, every morning after.