Connect Claude

Hobson runs a remote MCP server. Once Claude is connected, you can ask it things like "which customers are waiting longest?" or "summarise this week's refund requests" and it answers from your workspace.

The server is read-only. Claude can look things up and count things, but it can't send, change or delete anything. It never uses Hobson's own AI either: search is keyword and full-text search, and Claude does the thinking on your side.

claude.ai and Claude DesktopCopy Anchor Link

Add Hobson as a custom connector. You sign in with your Hobson account, so there's no key to copy.

1. In Claude, open Settings › Connectors and choose Add custom connector.
2. Name it Hobson and enter the URL https://api.hobson.support/v1/mcp.
3. Choose Connect. Claude opens Hobson in a new window. Sign in if you need to.
4. Pick one workspace for this connection. To use another workspace, disconnect and connect again.
5. Check the app name and the address it will send you back to, then approve the read access it asks for.

The connector then works in claude.ai, Claude Desktop and the Claude mobile apps. An owner or admin on a Team or Enterprise plan may need to add the connector for your organisation first.

What you approveCopy Anchor Link

Claude can never do more than your own role allows. If your role changes or you leave the workspace, the connection follows straight away.

DisconnectingCopy Anchor Link

Settings › Your account › Connected apps in Hobson lists every app connected to the workspace, with who connected it, what it can read and when it was last used. You can disconnect your own connections; owners and admins can disconnect anyone's. Disconnecting takes effect immediately.

Owners can switch Claude and other connected apps off for the whole workspace in the same place. While it's off, every connection stops working and nobody can connect a new one.

Claude CodeCopy Anchor Link

Claude Code can sign in the same way:

claude mcp add --transport http hobson https://api.hobson.support/v1/mcp

Then run /mcp, choose hobson and sign in.

Or use an API key, for scripts and shared setups:

1. Create an API key in Settings › Advanced › API keys. Viewer is enough.
2. Add the server with the key:

claude mcp add --transport http hobson https://api.hobson.support/v1/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

3. Start Claude Code and run /mcp to check that hobson is connected.

Claude Code stores the header in its own config. Treat that file like any other secret.

Other MCP clientsCopy Anchor Link

Clients that support OAuth for remote MCP servers sign in as above. Any client that supports streamable HTTP with a custom header can use an API key instead:

Available toolsCopy Anchor Link

Claude can summarise the workspace, list, search and read conversations, look up a customer (and their orders, when a shop is connected), search your knowledge, read reports and query metrics. Every tool is read-only.

• describeMetrics
• getBacklogReport
• getChatDeflectionReport
• getConversationContext
• getCustomerOrders
• getHobsonReport
• getQualityReport
• getReportTrends
• getResponseTimeReport
• getSlaReport
• getTeamReport
• getTicket
• getVolumeReport
• getWorkspaceSummary
• listTickets
• lookupOrder
• queryMetrics
• searchKnowledge
• searchTickets

Asking about numbersCopy Anchor Link

Two tools let Claude answer questions the fixed reports don't cover. describeMetrics lists what it can count, and queryMetrics counts it in the database and returns a small table, never message text.

• Metrics: conversations, messages, first reply time and resolution time (median, 90th percentile, average), replies per conversation, share handled by Hobson, auto-sent replies, drafts approved unchanged, CSAT, SLA breaches, reopened conversations and backlog.
• Group by up to two of: day, week, month, hour of day, weekday, channel, category, tag, status, assignee, team, priority and customer.
• Filter by channel, category, tag, status, assignee, team, priority, customer, and created or closed dates.
• Dates use your workspace's time zone unless Claude passes another. A query covers at most 366 days and returns at most 500 rows.

For example, "How quickly were tickets answered this week, by channel?" becomes:

{
  "metrics": ["first_reply_time.median", "first_reply_time.p90", "conversations"],
  "groupBy": ["channel"],
  "from": "2026-10-05",
  "to": "2026-10-11"
}

Times come back in seconds. To see the conversations behind a number, ask Claude to list or search them, for example "show me the slowest email ones".

What Claude seesCopy Anchor Link

• Flagged messages are withheld. When Hobson flags a message as a likely attempt to manipulate an AI assistant, Claude gets a short placeholder and a link instead of the text, and the conversation is marked injectionFlagged. Its subject, draft and the customer's name are withheld too. A person can read it in Hobson. Messages Hobson hasn't finished checking yet, and auto-replies and spam, are withheld the same way.
• Customer text is marked untrusted. Subjects, messages, names, attachment names, drafts, notes, knowledge passages and Shopify product text arrive inside <customer_content> tags. Hobson tells Claude to treat anything inside them as data, never as instructions. This makes manipulation less likely but can't rule it out, so check before letting Claude act on what it reads with other tools.
• Less personal data. Customer emails and phone numbers are masked (j•••@g•••.com), addresses are cut to town, postcode district and country, and payment details and order notes are left out. Hidden text and pasted card numbers are stripped.
• Results are capped. Each result is at most 40,000 characters. Long messages are cut at 4,000 characters, a conversation returns its newest 20 messages, and lists return up to 25 items. Pass nextCursor back as cursor to get the next page.

What Claude can't seeCopy Anchor Link

• Account lookup data. The plan, status, flags and fields your account lookup endpoint returns stay in Hobson. Claude can't call your endpoint or read cached answers.
• Linear. Claude can't see or raise Linear issues, and linked issues aren't included in conversations.
• Anything that changes the workspace. Settings, drafts waiting for approval, macros, automations, API keys and the audit log aren't available.

Open the conversation in Hobson to see these.

Good to knowCopy Anchor Link

• Every tool call is recorded in your workspace's audit log against the connection or API key.
• When a tool fails, Claude sees a short error like not_found: Conversation not found.
• Each call counts towards your plan's rate limits, whether it comes from a connection or an API key.