Connect an agent
Point Claude Code, Cursor, VS Code, Codex or a CI job at Gatehouse.
Every agent connects to the same URL: your Gatehouse address plus /mcp, for example https://gatehouse.yourcompany.com/mcp. You'll find it under Connect an agent in the console.
Agents sign in through the browser with OAuth 2.1 and PKCE. The person approves the agent once, and everything it does is recorded under their name.
Claude Code
claude mcp add --transport http gatehouse https://gatehouse.yourcompany.com/mcp
Then run /mcp in Claude Code, choose gatehouse and Authenticate. If Claude Code ever shows the server as disconnected, for example after Gatehouse restarts with new settings, run /mcp and choose Reconnect.
Cursor
Add Gatehouse to ~/.cursor/mcp.json:
{
"mcpServers": {
"gatehouse": { "url": "https://gatehouse.yourcompany.com/mcp" }
}
}
Open Cursor Settings → MCP. Gatehouse shows "Needs login". Click it and allow access in the browser.
VS Code
Add Gatehouse to .vscode/mcp.json:
{
"servers": {
"gatehouse": { "type": "http", "url": "https://gatehouse.yourcompany.com/mcp" }
}
}
Click Start above the entry and sign in when the browser opens.
CI jobs and scripts
Clients that can't open a browser use a personal token. Create one under Connect an agent → Tokens. Tokens expire after 30, 90 or 365 days. Use a separate token per job so you can revoke it on its own.
Send it as a header:
Authorization: Bearer gth_pat_…
For Claude Code without a browser:
claude mcp add --transport http gatehouse https://gatehouse.yourcompany.com/mcp \
--header "Authorization: Bearer gth_pat_…"
Revoking access
The Tokens list shows agent sign-ins and personal tokens. Revoke cuts that agent off immediately. Disabling a person revokes all their tokens at once.
What agents see
Agents see only the tools that are on and that some policy could allow for that person, plus three built-in tools for access requests. Gatehouse also tells agents, in its server instructions, not to retry blocked calls with different wording.