Quick answer: Connect Cursor to Gazebo to approve and record access to selected service credentials. OAuth and manual agent-token setup are different paths; successful retrieval returns the actual provider key to Cursor's runtime and potentially its host or model context.
Decide what access Cursor needs
Start with the provider credential, not the editor settings. Use a test environment and restricted provider permissions where available. A key stored in .env is not inherently full-account, and moving a broad key to Gazebo does not narrow its permissions.
Gazebo adds named retrieval identities, central access policies and retrieval history across services. That can be useful when several agents need different credentials. It does not execute Cursor's later Stripe, GitHub or Vercel requests.
For the maintained setup instructions, use the Cursor MCP guide. The following steps explain the decisions and verification behind that connection.
Option 1 — Connect with OAuth
Add a remote MCP server in Cursor using:
https://app.gazebohq.com/api/mcp
Use Cursor's sign-in flow. In Gazebo, authenticate, explicitly approve only the service keys this connection needs and choose an access deadline. Do not add a manual agent-token header to this OAuth setup.
The endpoint is shared; it is the authenticated connection and its policies that determine retrieval access. Registering an OAuth client does not itself grant access to keys.
Cursor Desktop has a tested Gazebo OAuth path. See the client compatibility guide for the recorded environment and evidence limits, rather than assuming every Cursor execution surface behaves identically.
Option 2 — Use a manual agent token
Create a named agent at Gazebo Agents, configure its retrieval policies and protect the resulting agent token. An account API token is not an MCP agent token.
Cursor's documented project configuration is .cursor/mcp.json; global configuration is ~/.cursor/mcp.json. For remote HTTP, a secret-free template can use:
{
"mcpServers": {
"gazebo": {
"url": "https://app.gazebohq.com/api/mcp",
"headers": {
"Authorization": "Bearer ${env:GAZEBO_AGENT_TOKEN}"
}
}
}
}
Set the variable in the environment inherited by Cursor. A desktop app launched from a launcher may not inherit your terminal's variables. Cursor's envFile option applies to stdio servers, not remote HTTP servers.
The template can be shared without a resolved secret. The actual token cannot: someone possessing it may retrieve credentials permitted by its policy. See MCP configuration security.
Verify identity before retrieving anything
Ask Cursor to call get_identity without fetching a credential. Check the returned identity, allowed services, declared-method policies and, for OAuth, approved key names.
Next, use a service containing a fictional test value. Substitute that service and key name in this illustrative call:
get_credential({ service: "your-test-service", key_name: "api_key", method: "GET" })
OAuth requires the approved key_name. Ask Cursor not to display the value, then inspect the matching retrieval event with list_audit_events. Not displaying a value does not mean Cursor's runtime or host never received it.
Finally, revoke the test grant or agent and confirm a fresh retrieval fails without reconnecting. Revocation does not invalidate a provider key already returned; use provider controls for that.
Keep project boundaries explicit
Use names such as cursor-billing-staging and avoid sharing one identity across unrelated work. A profile name is a useful label, not proof of a particular developer, task or session.
A project-specific configuration is not a sandbox. Global MCP entries, other enabled tools, readable files and inherited credentials may remain available. Telling Cursor “use project A” is not a security boundary; disable unrelated access and use appropriate runtime separation.
Troubleshoot without assuming the cause
Connection failure: check the client-specific configuration, endpoint and authentication mode. Distinguish an expired or revoked connection from a tool-level policy denial.
Retrieval denied: check the service, key name and declared method against the intended policy. Do not widen access merely to make an unexplained request succeed.
No retrieval event: first establish whether a retrieval was attempted. A connection problem, an identity-only check or no tool invocation can explain an empty log; it is not proof that Cursor found a key elsewhere.
Unexpected provider activity: compare retrieval events with Cursor/runtime and provider records. Retrieval time alone does not prove who performed a later operation or what data they accessed.
What this setup does—and does not—change
Gazebo manages which configured credentials Cursor can retrieve and records retrieval decisions. Its method parameter is release-policy context, not a way to make a write-capable key read-only. Tool-result handling, retention and model exposure depend on Cursor's deployment and settings; a separate display panel is not an isolation guarantee.
Use narrow provider permissions, trusted tools and appropriate execution controls alongside retrieval policy.
Last reviewed: 8 November 2026. Independent editorial; not affiliated with or endorsed by Cursor. Configuration references: Cursor's MCP documentation.