Manage RagChat from an AI client (MCP)
You can connect an AI client such as Claude to RagChat and let it work your support inbox for you: read conversations, draft and send replies, resolve threads, and (on standalone workspaces) write help center articles and manage automation rules. The connection uses MCP, an open standard for connecting AI clients to apps.
This is the reverse of Connect external tools to the AI (MCP), which lets the RagChat AI call other services while it chats with a customer.
Before you start
Connecting signs you in as a team member, so your workspace needs team access set up. If the MCP server section shows a warning about external access, your workspace isn't ready yet. Message us from the chat widget and we'll set it up.
There is no token to copy or keep secret. You approve the client in your browser, once.
Connect a client
-
Go to Settings → General and find the MCP server section.
-
Copy the command under Add to Claude Code and run it in your terminal. It looks like this:
claude mcp add --transport http ragchat https://app.ragchatai.com/mcpFor a client that asks for an address instead of a command, use the Server URL field above it.
-
The first call opens your browser with an approval screen. It shows which client is asking, the account you're signed in as, and the workspaces it will reach.
-
Click Approve.
The client's tools are ready right away. Some clients need a restart before new tools show up.
What a connected client can do
In the inbox, on every workspace:
- List and search conversations, and filter by unread or unresolved
- Read a full thread, including message attachments (screenshots, PDFs, and image-only messages with no text)
- Reply to a customer, or add an internal note only your team can see
- Resolve or reopen a conversation, and mark it read
- Turn AI mode on or off for a thread
- Edit conversation tags
- Edit or retract a message, block a contact, or delete a conversation
On standalone workspaces, it can also write, publish, and reorder help center articles, categories, and changelog entries. Help center tools don't apply to Shopify stores and will say so if a client tries.
To read many article bodies at once, call list-articles with include_content: true. Results are paginated (default 10 articles per page, max 25). Pass ids when you only need specific articles. Metadata-only listing (no bodies) can return up to 200 per page. Keep requesting the next page until page equals last_page. This avoids calling get-article once per article on large help centers.
On standalone workspaces, it can also manage automation rules:
- Read the full catalog of triggers, condition fields, operators, actions, tags, and limits (
describe-automation-options) - List rules, open one, create or update one, or delete one
- Inspect run history (
list-automation-runs) to see whether a rule fired for a contact and why not
Call describe-automation-options before creating a rule or changing its trigger, conditions, or actions. Invalid fields or operators are refused instead of silently never matching.
Creating a rule over MCP requires is_enabled: pass false unless you clearly want the rule live, and tell the client it is waiting to be enabled. Webhook bearer tokens are write-only (encrypted, never returned), same as in the UI. Automations tools don't apply to Shopify stores. See Set up automations.
Good to know
- One approval covers every workspace you belong to, including ones you join later. Each request names the workspace it acts on, so a client can't touch a workspace by accident.
- If you're removed from a workspace, the client loses access to it on its very next request.
- A reply sent this way is a real message to a real customer, exactly as if you typed it in the inbox. It takes the thread off the AI and reopens it, so the next customer message waits for a person. See Manage conversations in the inbox.
- Reading a thread doesn't clear its unread badge. Only marking it read does.
- A message with empty text and attachments is image-only, not empty. Open the attachment URLs before treating the message as blank.
- Deleting an automation also erases that rule's run history. Only do it when you mean to.
Review or revoke a connection
- Sign in at app.ragchatai.com.
- Click your name in the top right, then Connected apps.
- Each connection shows when it was approved and last used. Click Revoke to cut it off immediately.
Still stuck? Message us from the chat widget and we'll help.