The LogPulse MCP server is a remote Model Context Protocol server. It lets AI agents — Claude Code, Cursor, Codex, or any MCP-capable client — call typed LogPulse tools directly: search and investigate logs, query the risk-based SIEM, check service health, and walk the cross-domain entity graph. The agent reasons over structured results instead of scraping the dashboard or writing raw SQL.

The tools reuse the same in-process registry, LPQL engine, and namespace RBAC the product uses internally, so an agent never gets more access than the token holder. The surface is read-only by default; the one write tool creates disabled drafts for a human to review.

The server is exposed at a single endpoint over Streamable HTTP (stateless, JSON responses). All requests are POST and must be authenticated — an unauthenticated request is rejected with 401 before any tool is exposed.

POST https://api.logpulse.io/mcp

The endpoint and an authentication header are everything an MCP client needs; the client discovers the available tools, resources, and prompts automatically after it connects.

Two authentication methods resolve to the same identity (organization, user, scopes): a Personal Access Token as a bearer credential, or OAuth 2.1 for interactive sign-in.

Authorization: Bearer lpat_a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0c1d2e3f4

Scopes follow a resource:action convention. A token only sees the tools its scopes allow — a tool the token can't reach is invisible in tools/list and cannot be called. Follow least privilege and grant only what the agent needs.

Any MCP client that supports remote servers over HTTP connects the same way: the endpoint plus a bearer token. Below are the three most common clients.

claude mcp add logpulse --transport http \
  https://api.logpulse.io/mcp \
  --header "Authorization: Bearer lpat_your_token_here"

{
  "mcpServers": {
    "logpulse": {
      "url": "https://api.logpulse.io/mcp",
      "headers": {
        "Authorization": "Bearer lpat_your_token_here"
      }
    }
  }
}

Endpoint:  https://api.logpulse.io/mcp
Transport: Streamable HTTP
Auth:      Authorization: Bearer lpat_your_token_here

Tools are grouped by the scope that unlocks them. The catalog mirrors the depth your team uses in the dashboard.

Beyond callable tools, the server exposes MCP resources for read-grounding and prompts that steer the agent through a workflow.

Resources include an LPQL cheat sheet and a detection guide (available with any valid token), plus your data models and saved queries (gated by logs:read). A resource the token can't reach is hidden from resources/list.

Prompts include triage_notable, is_ip_malicious, onboard_connector, and service_health_check — each guides the agent through the right sequence of read-only tools for that task.

The MCP surface is governed by design:

Read-only by default. Only read tools are exposed. Writes are an explicit allowlist, each behind its own *:write scope; today that is just create_detection, which inserts a disabled rule.

Scoped & RBAC-aware. A token holds only the scopes you grant, and the token holder's namespace RBAC is applied to every query — an agent can never read across tenants or outside the user's teams.

Audited & rate-limited. Every tool call is written to the audit log (denied calls included) and rate-limited per token.

Usage is a detection. Each call is also emitted to an internal log stream and watched by built-in detections — abnormal call volume or an agent write raises a notable on your own engine.

Untrusted data stays data. Tool output is treated as data, not instructions, as a guard against prompt injection through your own logs. There is no auto-execute and no write without a human in the loop. Data stays EU-hosted and AI evaluation runs in the EU.

The endpoint is rate-limited to 120 requests/min per token. Exceeding it returns 429; back off and retry. Each successful call updates the token's "last used" timestamp in Settings, and every call is recorded in the organization's audit log with the user, scope, and whether it was a write.

Still stuck? See Personal Access Tokens for token errors, or contact us.