How to use Claude Code for customer support
Short answer. Add your support chat to Claude Code as an MCP server. The agent can then read conversations, reply to customers, save what it learned to the knowledge base, and turn what customers asked for into a brief it can build from. Replies and public changes can wait for your approval. On FanWake, Claude Code has made 825 tool calls to its support this way since 26 October 2026.
Most hard support questions are about how the product works, and the answer is in the code. A help desk can't read your code. A coding agent can, but it forgets everything between sessions and is only there while you are at the keyboard. Connecting the two over MCP gives you both: the support chat stays on your site and keeps what it learns, and the agent works the inbox with the code open.
I have run FanWake's support this way since late October 2026, with Afterchat as the support chat and Claude Code as the agent. Here is the setup, what the agent can do, and what it has done so far. The short version, with a demo, is on the Claude Code page, and every agent's setup is on the MCP server page.
What MCP is, in one paragraph
The Model Context Protocol is an open standard for connecting AI agents to outside tools. A service runs an MCP server that lists tools with descriptions, like "list conversations" or "reply to a customer". The agent reads the list, decides which tool fits the task, and calls it. Claude Code, Cursor, Codex and VS Code all speak it, so one server works with each of them.
Setting it up
Afterchat's MCP server lives at https://afterchat.chat/mcp and signs in with OAuth in the browser, so there is no API key in a config file. In Claude Code it is one command:
claude mcp add --transport http --scope user afterchat https://afterchat.chat/mcp
Restart Claude Code, run /mcp, pick afterchat and choose Authenticate. A browser opens where you pick the workspace and what the agent may do. Other agents:
# Codex
codex mcp add afterchat --url https://afterchat.chat/mcp
# Cursor: in ~/.cursor/mcp.json, under "mcpServers"
"afterchat": { "url": "https://afterchat.chat/mcp" }
# VS Code
code --add-mcp '{"name":"afterchat","type":"http","url":"https://afterchat.chat/mcp"}'
Then ask the agent to "check Afterchat status". It answers with the workspace, what it is allowed to do and what is waiting.
What the agent can do
The server has 22 tools. They fall into three groups, which are the three things a support chat leaves behind.
| Group | Tools | What it is for |
|---|---|---|
| Inbox | list and read conversations, reply, add a team note, close | Answering what was handed over, with the code open |
| Vault | search, list entries and gaps, write an entry or a guide, check what a change affects, organize topics, merge or archive, act on tidy-up suggestions | Keeping the knowledge the chat answers from correct and complete |
| Requests | list requests, get a request's brief, ask the people who requested it, ship | Building what customers asked for, then telling them it is done |
Eight slash commands package the common jobs. For example, /mcp__afterchat__inbox goes through what is waiting with the technical chats first, /mcp__afterchat__triage lays out the ranked requests for deciding what to build, /mcp__afterchat__build loads a request's brief, and /mcp__afterchat__vault builds or improves the vault from your code and docs.
What waits for a person
An agent that can message customers needs limits you can see. The ones I rely on:
- Replies go out at once or wait as drafts in the inbox, per workspace. Hold rules catch what should always get a person's eyes, like prices, refunds or legal terms, even when replies go out at once, and you can add rules of your own.
- Public changes to the help content wait on the Coding agent page until someone approves them. Right now 11 of the agent's changes are waiting there for me.
- Shipping a request, which emails everyone who asked for it, always waits for approval.
- Everything it does is logged, with which connection did it, so a change can be traced and undone.
What it has done on FanWake
| Since 26 October 2026 | Number |
|---|---|
| Tool calls from Claude Code | 825 |
| Checked entries in the vault | 164 |
| Entries that point at the code they were written from | 168 |
| Changes waiting for my approval | 11 |
| Open requests from customers | 8 |
Counted 16 November 2026. The entry counts overlap: an entry can point at code and be waiting for a check.
Most of the work went into the vault. On 31 October a Claude Code session in FanWake's repo rewrote it from the product's own code and docs: 152 entries in 10 topics, which I reviewed and approved as one batch. An entry about how a feature works points at the files it was written from, so when the code changes the agent can check which entries it affects. Since then the agent keeps the vault in shape over MCP. On 7 November it applied 28 merges of near duplicates and 4 moves between topics as one batch, which can be undone from the vault.
The part I haven't used much yet is the last step. None of FanWake's 8 requests has been built and shipped through the loop so far. The brief is there for each one, with the customers' own words and how many asked, but building from it is still ahead of me.
Other ways to do it
Several help desks now publish their own MCP servers, and there are community servers too, so check whether the tool you already use has one. If your support stack is your own, a small MCP server with a list and a reply tool is a weekend project. What to look for either way: sign-in with OAuth rather than a long-lived API key, a way to keep replies and public changes behind approval, and a log of what the agent did.
Questions
What is an MCP server for customer support?
It is a support tool exposing its inbox, knowledge base and requests through the Model Context Protocol, so an AI agent such as Claude Code, Cursor or Codex can read and act on them with tools, instead of you copying chats into a prompt.
Can Claude Code reply to customers?
Yes, through a reply tool, if you allow it. In Afterchat each workspace decides whether replies from an agent go out at once or wait as drafts for a person to approve, and hold rules can stop replies on topics like refunds.
Does it work with Cursor, Codex or VS Code?
Any agent that can add an HTTP MCP server can connect. Claude Code, Codex, Cursor and VS Code each have a one-line way to add it, and sign-in happens in the browser with OAuth, so there is no API key to paste.
Is it safe to give a coding agent access to support chats?
Treat it like giving a teammate access. The agent reads customer messages, which go to your model provider like any other prompt. Limit what it may do when you connect it, keep public changes behind approval, and check the audit log of what it did.