thruk-mcp

Model Context Protocol (MCP) server for Thruk — the unified web frontend for Naemon, Nagios, Icinga and Shinken.

Expose Thruk's REST API to MCP-compatible clients (Claude Desktop, Dust, LibreChat, OpenWebUI...) so that an LLM can query hosts/services, schedule downtimes, acknowledge problems, force rechecks and more in natural language.

Features

• Read: hosts, services, hostgroups, servicegroups, downtimes, comments, sites, aggregated stats, current problems

• Write: schedule/delete downtimes, acknowledge & remove acks, force rechecks

• Escape hatch: thruk_query tool to call any Thruk REST endpoint

• Multi-backend support (Thruk federated sites): pass backends="prod,dr" to any tool

• Transports: stdio (default) or Streamable-HTTP (--listen <port>, endpoint /mcp)

• Async httpx client with proper error handling and TLS verification

• Tested with pytest + respx, linted with ruff, packaged with hatchling

Related MCP server: vibeMK

Quick start

1. Configure

cp .env.example .env
$EDITOR .env   # set THRUK_BASE_URL and THRUK_API_KEY

An API key can be created from the Thruk user profile page (requires api_keys_enabled in thruk_local.conf) or via the REST API itself.

2a. Run with Docker

docker compose up -d
# MCP Streamable-HTTP endpoint: http://localhost:8001/mcp

2b. Run locally

pip install thruk-mcp        # or: pipx install thruk-mcp

# stdio mode (for Claude Desktop, LibreChat, etc.)
thruk-mcp

# Streamable-HTTP mode — endpoint http://localhost:8001/mcp
thruk-mcp --listen 8001
# equivalently: thruk-mcp --transport streamable-http --listen 8001

# Behind a load balancer / multiple replicas, drop per-session state
# (no sticky routing required):
thruk-mcp --listen 8001 --stateless --json-response

# Multi-tenant: each request brings its own Thruk credentials via headers
# (no fixed THRUK_API_KEY at boot). Requires --stateless; serve over TLS.
thruk-mcp --listen 8001 --stateless --header-auth

For local development of the project itself, see CONTRIBUTING.md.

3. Wire it to an MCP client

Claude Desktop (~/.config/Claude/claude_desktop_config.json or macOS equivalent):

{
  "mcpServers": {
    "thruk": {
      "command": "thruk-mcp",
      "env": {
        "THRUK_BASE_URL": "https://monitor.example.com/thruk",
        "THRUK_API_KEY": "xxxxxxxx"
      }
    }
  }
}

4. Use with the Docker MCP Gateway

The image at ghcr.io/k9fr4n/thruk-mcp:latest defaults to stdio transport, so it can be spawned natively by the gateway.

Option A — Private local catalog

# 1. Create your private catalog
docker mcp catalog create thruk-private

# 2. Register this server (catalog/server.yaml ships with the repo)
docker mcp catalog add thruk-private thruk-mcp ./catalog/server.yaml

# 3. Configure credentials & enable
docker mcp secret set thruk-mcp.api_key=YOUR_KEY
docker mcp config write thruk-mcp.base_url=https://monitor.example.com/thruk
docker mcp server enable thruk-mcp

# 4. Run the gateway with your catalog
docker mcp gateway run --catalog thruk-private

Then point any MCP client (Claude Desktop, VS Code, Cursor, ...) at the gateway as documented here.

Option B — Submit upstream

catalog/server.yaml, catalog/tools.json and catalog/readme.md follow the docker/mcp-registry schema and can be submitted to the official Docker MCP Catalog via PR.

What's exposed

65 MCP Tools

Read — state thruk_list_hosts, thruk_get_host, thruk_list_services, thruk_get_service, thruk_list_hostgroups, thruk_list_servicegroups, thruk_list_contacts, thruk_get_contact, thruk_problems, thruk_stats, thruk_totals (compact 16-field host+service totals, faster than thruk_stats), thruk_sites.

Read — history & comments thruk_list_logs, thruk_list_alerts, thruk_list_notifications, thruk_notification_summary (notifications grouped by contact/host/service/state/command), thruk_recent_events, thruk_list_comments, thruk_list_downtimes, thruk_get_downtime, thruk_state_at (reconstruct the parc state at a past instant from /logs — a post-mortem snapshot), thruk_state_diff (what changed between two past instants t1 → t2, replayed from /logs).

Read — noise & flap analysis thruk_top_noisy_hosts (hosts ranked by alert count over a window), thruk_top_noisy_services (services ranked by alert count), thruk_flap_summary (hosts/services ranked by state transition count).

Read — problem intelligence thruk_oldest_problems (unhandled problems sorted by age, oldest first), thruk_unacked_critical (CRITICAL/DOWN not acknowledged for > N minutes), thruk_stale_acks (acknowledgements older than N days — forgotten problems), thruk_problem_counts (flat aggregate of unhealthy-state counts, filterable by hostgroup, custom vars or any structured filter — replaces the former thruk_problems_by_hostgroup), thruk_stale_checks (surface checks that stopped running — the dangerous "false green"), thruk_backend_health (per-site supervision-backend health: latency, replication lag, blind spots), thruk_worker_health (distinguish a real outage from a mod-gearman supervision blind spot).

Read — analytics thruk_alert_heatmap (alert counts bucketed by time, useful for spotting recurring patterns), thruk_notification_heatmap (notification counts bucketed by time — spot mail/paging storms), thruk_concurrent_failures (windows where multiple hosts failed simultaneously), thruk_recurring_problems (hosts/services generating repeated alerts over a window), thruk_root_cause (collapse a DOWN/UNREACHABLE storm into its root cause(s) via parent topology), thruk_unreachable_vs_down (split a host outage window into DOWN cause vs UNREACHABLE consequence).

Read — availability / SLA thruk_host_availability (uptime % for a single host — time_up_percent, time_down_percent, time_unreachable_percent and scheduled equivalents), thruk_service_availability (ok/warning/critical/unknown % for a single service), thruk_hostgroup_availability (availability for all hosts or services in a hostgroup, sorted worst-first; type = hosts | services | both), thruk_hostgroup_availability_summary (one aggregated rollup instead of one row per host — time-weighted availability_percent, worst/best, below_threshold count, state distribution; ideal for incident/SLA reports on large groups). All accept since/until (Thruk relative or ISO) or a timeperiod shortcut (lastmonth, thismonth, last24hours, lastweek, …). thruk_reliability_report (per host/service reliability metrics — MTTR / MTBF / incident counts — derived from the log over a window). thruk_incident_timeline (ordered event chronology — the post-mortem "déroulé" — for a host, service or hostgroup: every state change, notification, downtime, flap and acknowledgement in time order, plus an incident/MTTR summary; a scoping filter is required).

Read — performance data thruk_get_perfdata (fetch and parse performance data for a single host or service), thruk_perfdata_snapshot (parsed perfdata for every service matching a filter, in one call), thruk_perfdata_near_threshold (metrics within within_percent % of breaching their warn/crit range — early-warning signal before an alert fires).

Write — downtime management thruk_schedule_downtime (host/service), thruk_schedule_host_services_downtime (all services of a host), thruk_schedule_propagated_host_downtime (parent+children), thruk_schedule_hostgroup_downtime, thruk_schedule_servicegroup_downtime, thruk_delete_downtime, thruk_delete_active_downtimes, thruk_delete_downtimes_by_filter.

Write — problem handling thruk_acknowledge, thruk_bulk_acknowledge (acknowledge multiple hosts/services in one call), thruk_remove_acknowledgement, thruk_recheck, thruk_add_comment, thruk_delete_comment, thruk_checks (enable/disable active checks for a host or service), thruk_notifications (enable/disable host or service notifications, with optional cascade to all services of a host).

Escape hatches thruk_query (raw call to any REST endpoint), thruk_run_background_query (long-running endpoint via Thruk's ?background=1 mechanism with automatic job polling).

All list-style tools share a consistent limit / offset / sort / columns contract. By default they return a tight subset of columns (~10 fields per row) to keep LLM token consumption low. Pass columns="" to opt out and receive every column the Thruk row contains.

5 MCP Resources

URI templates that MCP clients with a resource browser (Claude Desktop, VS Code, ...) can "open" like files:

3 MCP Prompts

Pre-canned workflows the user can invoke as a slash-command in the MCP client UI:

Robustness

• Connection retries — httpx.AsyncHTTPTransport(retries=3) handles DNS failures, connection refusals, TLS handshakes.

• HTTP retries with backoff — 5xx and 429 responses are retried up to 3 times with exponential backoff + jitter (cap 5 s).

• Opt-in TTL cache — slow-moving endpoints (/sites, /processinfo, /hosts/stats, /services/stats, /contacts, /timeperiods, ...) are cached in-process for 15 s. Any tool can request caching via cache_ttl= on the underlying client. This absorbs the burst of identical calls an LLM agent typically issues across a multi-tool turn.

• Pagination helper — ThrukClient.get_all() is an async generator that iterates pages of 500 rows up to a configurable hard limit (default 50 000), so internal callers can scan entire backends without manual offset math.

• Long-running queries — the thruk_run_background_query tool wraps Thruk's ?background=1 flow and polls /thruk/jobs/<id>/output until the job completes (5 min default timeout).

Environment variables

Connection

Security / multi-tenant (v0.6)

Security

• Read-only mode — set THRUK_READ_ONLY=true to remove every write tool (thruk_acknowledge, thruk_schedule_*_downtime, thruk_recheck, thruk_delete_*, thruk_run_background_query) from the MCP server. The LLM literally cannot mutate monitoring state. Use this for general-purpose agents that should only observe.

• Tool allowlist — THRUK_ENABLED_TOOLS=thruk_list_*,thruk_problems,thruk_stats restricts the exposed surface to the listed tools (fnmatch wildcards supported). Useful when fronting multiple LLM clients with the same gateway but different scopes.

• Audit log — every write tool invocation emits one JSON line on thruk_mcp.audit (stderr by default):

  {"ts":"2026-05-17T22:00:00+00:00","tool":"thruk_acknowledge","user":"alice",
   "args":{"host":"srv01","comment":"investigating"},"target":"srv01","status":"ok"}

  Disable with THRUK_AUDIT_LOG=false. Sensitive keys (api_key, password, token) are redacted as *** before logging.

• Rate limit — THRUK_MAX_CONCURRENT=8 caps in-flight HTTP requests with an asyncio.Semaphore. Combined with the v0.3 TTL cache, this protects the Thruk core from an LLM that loops on tools or chains them aggressively.

• Transport-level HTTP auth — gate the Streamable-HTTP /mcp endpoint itself, independently of the Thruk credentials a request carries (no effect on stdio). Set MCP_HTTP_TOKEN=<secret> to require an Authorization: Bearer <token> header (constant-time compare; 401 + WWW-Authenticate: Bearer otherwise). HTTP serving fails closed — --listen / --transport streamable-http refuses to start unless MCP_HTTP_TOKEN is set or MCP_HTTP_ALLOW_UNAUTHENTICATED=true is given (explicit opt-out for proxy-fronted deploys). MCP_HTTP_ALLOWED_HOSTS enforces a Host allowlist (anti-DNS-rebinding, defaults to loopback). The chain is TrustedHost → Bearer → HeaderAuth → /mcp, so the bearer gate composes with header-auth multi-tenant mode below.

• Header-auth multi-tenant — run thruk-mcp --listen 8001 --stateless --header-auth (or THRUK_HTTP_HEADER_AUTH=1) to serve many users from one process, each with their own Thruk credentials supplied per request via headers:

  Header	Maps to	Required
  X-Thruk-Auth-Key	api_key	yes (else 401)
  X-Thruk-Base-Url	base_url	no (falls back to THRUK_BASE_URL)
  X-Thruk-Auth-User	auth_user	no
  X-Thruk-Backends	default_backends (CSV)	no

  The server boots without THRUK_API_KEY. Only credential/endpoint fields come from headers — THRUK_READ_ONLY, THRUK_ENABLED_TOOLS and THRUK_AUDIT_LOG remain server-owned, so a tenant cannot grant itself write access or silence the audit log (which still attributes each call to the tenant's auth_user). Per-tenant HTTP clients are pooled in a bounded LRU cache. The API key travels in a header, so serve only over TLS (terminate TLS in front, or behind a trusted reverse proxy). Requires --stateless.

Development

pip install -e ".[dev]"
pre-commit install                              # one-time setup of git hooks

ruff check src tests && ruff format src tests   # lint + format
mypy src                                        # type-check
pytest -v --cov=thruk_mcp --cov-fail-under=80   # tests with coverage gate

Conventions:

• Conventional Commits (feat:, fix:, chore:, docs:, refactor:, test:).

• No direct push to main: branch → PR → squash merge.

• Any new tool must come with a respx-mocked unit test in tests/test_tools.py; regenerate catalog/tools.json (Docker MCP Registry contract) with python scripts/gen_tools_json.py — it is generated from the live registry, not hand-edited, and CI enforces it via --check.

• CI gate: ruff, ruff format --check, mypy, pytest with 80 % coverage minimum.

References

• Thruk REST API: https://www.thruk.org/documentation/rest.html

• Thruk REST commands: https://www.thruk.org/documentation/rest_commands.html

• MCP spec: https://spec.modelcontextprotocol.io/

• Inspired by: https://github.com/lausser/omd-mcp (initial proof-of-concept)

Project docs

• CHANGELOG.md — what changed in each release.

• UPGRADING.md — per-version migration notes.

• SUPPORT.md — supported Python / Thruk / MCP-client versions, security policy, release cadence.

• CONTRIBUTING.md — dev setup, PR conventions, tool / env-var contribution checklists.

License

MIT — see LICENSE.