MCP server: connect an AI client
How Claude, Cursor or any MCP client connects to PulseTrace over OAuth — registering the client, the access policy and approvals, and the full list of tools it can call.
PulseTrace runs an MCP server so an AI client can list, create and inspect monitors, read incidents and intelligence, and manage notification rules — inside the permissions you grant it. Access is per scope, every tool call is logged, and destructive tools can require a human approval.
How a client connects
- Endpoint:
POST <mcp base url>/mcp— the base URL is shown on MCP → Registered clients in the app. JSON-RPC 2.0 over HTTP, protocol version2025-06-18, supportinginitialize,tools/list,tools/call,resources/listandresources/read. There is no SSE transport and no prompts capability. - Authentication: an OAuth 2.1 access token as
Authorization: Bearer …. API keys are not accepted by the MCP endpoint. The server publishes/.well-known/oauth-protected-resource,/.well-known/oauth-authorization-serverand/.well-known/openid-configuration, and the client walks the standard authorization-code + PKCE flow through/oauth/authorizeand/oauth/token. - Dynamic client registration is not offered: register the client first (below) so PulseTrace knows its redirect URIs.
Register the client
- 1Go to MCP → Registered clients and add the client: a name, a client type of
public(PKCE only, no secret — desktop apps and editors) orconfidential(server-side, gets a secret shown once), and its redirect URIs, one per line, matched exactly. - 2In the AI client, add PulseTrace as a remote MCP server at the endpoint above and start the OAuth flow; sign in as a member of the organization and approve the requested scopes.
- 3The grant appears under MCP → Connected clients, where it can be revoked at any time. Sessions and every tool call are listed under MCP → Activity.
Exact configuration fields differ per client (some take a URL only, others a command that launches a bridge). Because PulseTrace uses standard OAuth discovery, a client that supports remote MCP servers with OAuth needs only the endpoint URL.
Access policy
MCP → Policy governs every client in the organization:
- MCP enabled; allow external AI clients; require user authorization for new connections; require MFA before authorizing.
- Default access level:
read_only,monitoring_operator,full_managementorcustom; allow write operations. - Require confirmation for destructive operations —
disable_monitoranddelete_monitorthen open a pending approval that an owner or admin must decide within 24 hours. - Who can authorize (
all,admins_only,selected), client restrictions (anyorapprovedonly), and IP restrictions as a CIDR list.
Tools
| Tool | Scope |
|---|---|
list_monitors, get_monitor, get_monitor_status, get_monitor_results, list_executions, get_execution, get_execution_steps, get_step_history | monitors:read / results:read |
create_monitor, update_monitor, enable_monitor, trigger_execution, run_monitor_check, parse_openapi_spec | monitors:create / monitors:update / monitors:enable |
disable_monitor, delete_monitor | monitors:disable / monitors:delete — approval-gated by policy |
get_monitor_dependencies, set_monitor_dependencies | monitors:read / monitors:update |
list_incidents, get_incident, list_anomalies, get_anomaly | incidents:read / anomalies:read |
get_intelligence, get_monitor_recommendations, list_discovered_hostnames, list_domain_verifications | intelligence:read |
get_security_posture, list_posture_scores, get_posture_history, get_pci_report | intelligence:read |
list_notification_rules, create_notification_rule, update_notification_rule, delete_notification_rule | notification_rules:* |
list_monitor_credentials, create_monitor_credential, delete_monitor_credential | credential management (owner/admin) |
list_regions, get_organization | agents:read / organization:read |
MCP access is part of API access and starts on the Team plan. Audit rows for MCP are kept for 365 days; approval requests for 30 days after they expire.