Connect an AI agent

This guide connects Claude Code (or any MCP client) to CHouse UI with a token that can only read, then shows how to watch what it does. Two people may be involved: an administrator with agents:manage, and you.

1. Administrator: turn MCP on

1. AI Governance › MCP → switch MCP server on.
2. Check the endpoint shown, e.g. https://chouse.corp/mcp. If agents reach CHouse UI on another address, set Public address.
3. Leave the tool defaults — read tools on, everything else off — for now.

2. You: create a narrow token

Preferences › Personal access tokens › New token:

• Name: claude-code-laptop
• Expiry: 30 days
• Scopes: only what the agent needs, e.g. database:view, table:view, table:select, query:execute, observe:view

Copy the token and keep it in your environment, never in a config file:

export CH_HOUSE_PAT=ch_pat_…

3. Register the server with the client

claude mcp add --transport http chouse https://chouse.corp/mcp \
  --header "Authorization: Bearer $CH_HOUSE_PAT"

chouse mcp config claude-code    # or codex, cursor, vscode, opencode

Other clients: MCP › Connect a client.

4. Check what the agent can do

chouse mcp status    # on/off, endpoint, and the tools your token gets
chouse mcp tools

Ask the agent something real — "which tables in shop are stale, and what feeds them?". It should call get_dataset_health and get_lineage, and read get_table_context before writing SQL.

5. Watch it

AI Governance › Sessions shows the session: every tool call, bytes read, budget used and any health notices it received. Administrators can set limits in AI Governance › Policies — max read per query, a daily budget, partition filters on large tables, and what to do with tables that have incidents — or pause all agent access at once.

6. Give it more, carefully

To let the agent kill queries or run DDL, an administrator turns those tools on in AI Governance › MCP (with a confirmation), and your token needs the matching permissions. Make destructive calls ask you first in the client — see Human approval per client.