Cline MCP guide: triage customer support without leaving VS Code

Your code lives in VS Code. Cline sits in the sidebar with your Anthropic key loaded, ready to chain twenty tool calls without complaining. Meanwhile your support inbox is a browser tab you have not opened since yesterday, with nine conversations waiting, three of which need someone who can actually read the code to answer. The context switch is the real cost: open the tab, read the thread, dig through the repo, write the reply, lose your place. This guide removes the tab. You will connect Cline to your muro inbox over MCP, so support triage happens in the editor, with the codebase in the same context window.

Setup takes about ten minutes: one API key, one entry in cline_mcp_settings.json, one test prompt. Then we get to the interesting part, which is what an agent that can chain tool calls does with an inbox.

Why Cline is a good fit for support triage

Plenty of MCP clients can read a support inbox. Cline, the open-source VS Code agent, has three properties that make it unusually good at acting on one.

• →It chains tool calls. Ask a one-shot chat integration to triage your inbox and you get a summary. Ask Cline and it lists conversations, opens each thread, groups them, and drafts replies, one tool call after another, until the job is actually done.
• →Approval is built into the loop. By default Cline asks before every tool call. You can auto-approve the harmless read operations and keep a manual click on anything that sends words to a customer.
• →It sits next to your code. When a user reports that your webhook retries are failing, Cline can read the conversation and open the handler file in the same run. No browser-based tool can do that.

That last point is the whole pitch. Support questions about technical products are usually code questions wearing a trench coat. The agent that answers them should be able to see the code.

What you need

• →Cline installed in VS Code. It is free and open source; install it from the extension marketplace.
• →A muro account with at least one site. The trial is 14 days, no card.
• →A muro API key, which we will create in the next step.
• →Node 18 or newer, because the connection runs through npx mcp-remote.

Step 1: create a scoped API key

In your muro dashboard, head to the API keys section of your settings and create a new key. Two decisions matter here, and they are the difference between a calm setup and a nervous one.

• →Read vs write scope. A read-scoped key lets Cline list conversations, open full threads, and search. A write-scoped key also lets it send replies and add internal notes. Start with read-only. Create a second key with write scope once you have watched the agent work for a few days.
• →Per-site scope. muro keys can be scoped to a single site. If you run three projects, a key scoped to project A physically cannot read project B's conversations. One key per site, per machine, is the boring correct answer.

The key looks like mr_live_... and is shown once. Treat it like any other secret: it goes in the config below, never in a repo.

Step 2: add muro to cline_mcp_settings.json

Cline manages its MCP servers in a file called cline_mcp_settings.json. The fastest way to open it: click the MCP Servers icon in Cline's top toolbar, switch to the Installed tab, and click Configure MCP Servers. The file opens right in the editor. Add muro under mcpServers:

json{
  "mcpServers": {
    "muro": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://muro.chat/api/mcp",
        "--header",
        "Authorization: Bearer mr_live_YOUR_KEY"
      ]
    }
  }
}

That is the entire integration. muro's MCP server is hosted at https://muro.chat/api/mcp, and mcp-remote bridges it to the stdio transport Cline speaks. Nothing runs locally except the bridge, so there is no server process to install, update, or babysit. Save the file and Cline reloads it; you should see muro appear in the MCP panel with a green light and a list of tools for conversations, messages, and replies.

Step 3: prove it works with read-only prompts

Open a new Cline task and ask something harmless:

• →"List my open muro conversations from the last 24 hours."
• →"Show me the full thread of the conversation about webhook retries."
• →"Which open conversations mention billing or invoices?"

Cline will ask permission for each tool call the first time. Approve them and watch what it does: the sequence of calls it makes is the best documentation of what the muro server exposes. (The underlying REST surface, if you ever want to script against it directly, is in the API docs.)

Once you trust the read tools, open Cline's auto-approve settings and allow them. This is what makes the next section feel like magic instead of a click marathon.

The payoff: multi-step triage that actually executes

Here is the prompt that justifies the setup. Type this into Cline at 9am:

Because Cline chains tool calls, this is not a summary request. It executes:

1. 01List conversations opened or updated since yesterday evening.
2. 02Open each thread and read it in full, one call at a time.
3. 03Group them: say, two reports about the same broken webhook, one feature request, one pricing question, one refund demand.
4. 04Flag the refund thread and skip drafting it, because you told it to.
5. 05For the webhook reports, read the relevant handler in your repo and draft technically accurate replies.
6. 06Present all the drafts and stop.

You read the drafts, edit one, then say: "Send the replies for the webhook conversations and the pricing one." With a write-scoped key configured, each send is a separate tool call that pops up for approval. You click approve four times and your queue is empty before your coffee is cold.

One Cline-specific tip: use Plan mode for the first pass. Ask it to plan the triage, check which conversations it intends to touch and roughly what it intends to write, then flip to Act mode to execute. For an agent that talks to customers, a visible plan beats a fast surprise.

Keep the human on the send button

An agent with write access to customer conversations deserves a little paranoia. These rules have held up:

• →Never auto-approve write tools. Reads, fine. But every reply that reaches a customer should cost you one deliberate click. The failure mode is not the agent going rogue; it is a plausible-sounding draft with one wrong fact going out unread.
• →Drafts before sends. Phrase your prompts as "draft replies" and send explicitly afterwards. It keeps you in the editor role, which is the right role.
• →Escalation rules live in the prompt. Refunds, account deletions, security reports, and anyone who is clearly angry go to a human. Put that in your standing prompt, the same way muro's own AI auto-reply is built to hand off to a human when it is unsure rather than guess.
• →Narrow keys. Read-only where possible, per-site always, revoked the moment a machine changes hands.

Close the loop: triage feeds deflection

The MCP connection is the founder-facing half of muro. Your customers still meet the other half: a small async widget on your site (about 6 KB), AI auto-reply grounded in your published help articles and product context, email forwarding so support@ lands in the same inbox with reply-by-email. Cline is the power tool for whatever that first line cannot handle.

And triage output is deflection input. If Cline drafted the same webhook explanation three times this week, your docs have a hole. Ask it to turn those three threads into a help article draft, publish it, then import the article URL in muro so the AI answers that question on the widget automatically. Next week that question never reaches your queue at all. Triage in the editor, publish the answer, train the AI on it: that loop is how a solo founder's inbox stays boring.

If you want the whole loop running today: create a site in muro (14-day trial, no card required), paste the widget snippet on your site, import your docs URL so the AI has something to stand on, and drop the JSON block above into cline_mcp_settings.json. Tomorrow at 9am, ask Cline what came in overnight, and see how it feels to finish support before standup.