MCP overview
Documentation index (llms.txt)View as Markdown
Gazebo exposes a hosted MCP (Model Context Protocol) server for credential retrieval. Use a remote Streamable HTTP client with MCP OAuth and Client ID Metadata Documents (CIMD) or Dynamic Client Registration (DCR), or with a custom bearer-token header. Client configuration and authentication support vary.
Endpoint
https://app.gazebohq.com/api/mcp
With MCP OAuth, sign in to Gazebo, explicitly select the service keys to share and choose an access deadline. The connection uses the existing agent policies and grant controls; it does not grant account administration access. Disconnect it from Settings → Developer → Consent clients → Active Grants, or revoke its linked agent. OAuth clients can use CIMD or dynamically register a public client; neither registration method grants credential access by itself.
For manual bearer authentication, create an agent and its token at Agents and pass it in the Authorization header. Account API tokens and browser sessions cannot authenticate MCP requests.
For an OAuth connection, get_credential must specify the approved key_name. Omission does not select a default key or grant access to other keys. The returned provider key remains visible to the client runtime and its host; revoking the connection only stops future retrieval.
Available tools
Tool metadata and permissions
Each tool advertises a human-readable title and standard MCP behavior hints:
| Tool | Display title | Read-only | Destructive | Idempotent | Open-world |
|---|---|---|---|---|---|
get_identity | View connection identity and access | Yes | No | Yes | No |
get_credential | Retrieve a service credential | No | No | No | No |
list_audit_events | List credential-access events | Yes | No | Yes | No |
Credential retrieval is not marked read-only: valid attempts create audit events, and successful retrieval releases a sensitive value to the runtime and its host. It is not destructive because it does not modify or delete provider data. Open-world is false because these tools operate on Gazebo's stored credentials and account records, not arbitrary provider endpoints.
These hints describe behavior; they do not grant access, enforce client confirmation or guarantee safe handling of returned keys. Gazebo's authentication and retrieval policies still apply. Client permission controls vary.
get_identity
Returns the calling agent's name and service/method access policy. Call this first to confirm which services are accessible before fetching credentials. No credentials or credential-access audit events are produced; ordinary operational request logging may still occur.
get_identity()
→ {
status: "ok",
agent_id: "...",
name: "cursor-dev",
accessible_services: [
{ service: "stripe", allowed_methods: ["GET", "POST"] },
{ service: "github", allowed_methods: ["GET"] }
]
}
get_credential
Retrieves the credential for a named service. It requires service and method. The declared method is checked against the credential-release policy; the tool does not execute that HTTP method or any provider request. OAuth requests also require an explicitly approved key_name. Only manual agent-token requests may omit key_name to select the service's primary active credential. Valid credential-access attempts (including policy denials) are logged; authentication failures and malformed arguments are not credential-access events. The actual stored provider key is delivered to the MCP client/agent; avoid persisting it.
See the Get credential reference for full documentation.
list_audit_events
Returns recent credential-access attempts for the calling agent. Use limit from 1 through 50, and optionally filter with service and paginate with cursor. Every event includes a stable event_id for correlation.
list_audit_events({ limit: 10, service: "stripe" })
→ {
status: "ok",
count: 2,
events: [
{ event_id: "...", service: "stripe", method: "POST", outcome: "success", timestamp: "..." },
{ event_id: "...", service: "stripe", method: "GET", outcome: "denied", timestamp: "..." }
]
}
Recommended usage pattern
1. get_identity() → confirm which services are accessible
2. get_credential(...) → fetch the credential you need
3. [use the credential] → make your API call; do not persist the value
4. list_audit_events(...) → optionally verify the access was logged
Client setup
Choose your client below. Start with get_identity to confirm the connection before retrieving a credential.
For Claude Desktop, use the remote custom connector flow described in Claude's official guide, not Claude Code's configuration. Add https://app.gazebohq.com/api/mcp, select Sign in now, Use Claude's published identity, and Streamable HTTP. Complete Gazebo's sign-in and approve only the keys this connection needs. Claude's connector controls and availability depend on your account. Its remote connector runs through Anthropic's infrastructure, which can receive returned credentials; this is not a local-only connection. Do not select No sign-in for the OAuth flow.
Tested clients
Tested paths include Cursor Desktop, Claude Chat, Claude Desktop, ChatGPT Business on web, Claude Code's desktop Code session, Devin hosted, Codex CLI and the OpenAI Responses API. The latest recorded client verification is 8 November 2026; this is not a claim that every client was retested on that date.
| Client or runtime | Recorded test scope |
|---|---|
| Cursor Desktop | macOS OAuth connection, tool discovery, identity, reported fictional-key retrieval and rejection after revocation. |
| Claude Chat | Normal Chat OAuth flow on 7 November: identity, approved fictional-key retrieval, matching audit retrieval and rejection after grant revocation. |
| ChatGPT Business — web | User-tested on 8 November: OAuth connection, identity, fictional-key retrieval and matching audit lookup. After revocation, a fresh retrieval prompted reconnection. |
| Claude Desktop | macOS remote connector OAuth flow; the earlier Desktop run covered discovery, identity, named retrieval and rejection after revocation. |
| Claude Code — desktop Code session | macOS OAuth flow on 6 November: discovery, identity, fictional-key retrieval, matching audit retrieval and rejection after revocation. |
| Devin hosted | Personal OAuth flow executed on Devin's servers, including identity, fictional-key retrieval, matching audit retrieval and rejection after revocation. |
| Codex CLI | Native app-server MCP tool calls using a scoped agent token; not a model-driven, IDE or OAuth test. |
| OpenAI Responses API | A separate model-driven test covering credential retrieval, denials, audit pagination and revoked-token failure. |
Claude Code's terminal/CLI run verified connection health and rejection after revocation, not tool execution. Other ChatGPT plans, native mobile clients and token refresh were not verified by the Business web test; public-directory eligibility is separate. Legacy Windsurf/Cascade remains unverified. A successful test in one runtime or authentication mode does not verify another.
See the compatibility guide for tested authentication methods, environments and detailed results.