Es Error Lens

An MCP server that gives LLM agents eyes on your Elasticsearch logs. Point it at a cluster holding ECS-format logs and any MCP client (Claude Desktop, Claude Code, or your own agent) can search errors, detect recurring patterns, analyze error-rate trends, and pull full trace context — the queries an engineer runs by hand during triage, exposed as tools.

Tool	What it answers
search_errors	"What's failing right now?" — filtered by window, service, level
get_error_patterns	"Is this systemic or a one-off?" — message aggregation with occurrence counts
analyze_error_trend	"Is it getting worse?" — time-series histogram, peak detection, trend verdict
get_error_context	"What led up to this?" — every log sharing the error's trace ID, in order
compare_errors	"Are these two failures related?" — attribute + fuzzy-message comparison
health_check	"Can I even reach the cluster?"

Elasticsearch is spoken to over its plain REST API via httpx — no Elasticsearch client dependency, and the HTTP layer is transport-injectable, so the entire server runs offline against the bundled FakeElasticsearch for tests and the demo.

Demo (offline, no cluster, no API keys)

git clone https://github.com/TerraCo89/es-error-lens && cd es-error-lens
python -m venv .venv && source .venv/bin/activate   # .venv\Scripts\activate on Windows
pip install -e ".[dev]"

pytest              # 22 tests, all offline, < 1 second
es-error-lens-demo  # synthetic incident walk-through

The demo mounts FakeElasticsearch with a deterministic 60-document ECS corpus simulating a small incident — a steady payment-provider failure, an escalating search-timeout problem, and background noise — then triages it:

[get_error_patterns] 2 recurring patterns in 57 errors:
   42x  [TimeoutError] Search query timed out after 30s  (search-api)
   13x  [UpstreamHTTPError] Payment provider returned 502 Bad Gateway  (checkout-api)
  error types: TimeoutError=42, UpstreamHTTPError=13, TemplateError=1, ConnectionResetError=1

[analyze_error_trend] search-api: 42 errors, trend=INCREASING
  peak: 8 errors at 2026-07-10T06:00:00Z

[get_error_context] trace-checkout-7f3a -> [UpstreamHTTPError] Payment provider returned 502 Bad Gateway
  3 related logs on the same trace:
    info  POST /checkout started
    info  Cart validated: 3 items
    warn  Payment provider latency 4100ms exceeds SLO

Demo result: OK

Using it against a real cluster

export ELASTICSEARCH_URL=http://localhost:9200   # default
export ES_INDEX_PATTERN=logs-*                   # default

es-error-lens                                    # stdio, for Claude Desktop / Claude Code
es-error-lens --transport streamable-http --port 8080   # HTTP clients

Claude Desktop / Claude Code config:

{
  "mcpServers": {
    "es-error-lens": {
      "command": "es-error-lens",
      "env": { "ELASTICSEARCH_URL": "http://localhost:9200" }
    }
  }
}

Expected document shape — standard ECS fields, the ones Filebeat and the ECS logging libraries emit by default: @timestamp, log.level, message, service.name, error.type, error.stack_trace, trace.id, host.name. Anything missing degrades gracefully (fields come back null).

Testing without a cluster

FakeElasticsearch (also exported) implements just enough of the _search API for this server: bool queries with term/range clauses, sorting, terms aggregations with top_hits sub-aggregations, and date_histogram with fixed intervals. Mount it in your own tests:

from es_error_lens import FakeElasticsearch, set_transport, search_errors

set_transport(FakeElasticsearch(my_ecs_docs).transport())
result = await search_errors(time_range="1h", service="checkout-api")

The test suite runs entirely through it — deterministic, offline, fast — and a hygiene test enforces that all bundled demo data stays synthetic (hosts under .example, no real domains or addresses).

Provenance

Extracted from a personal observability platform where it fronts the Elasticsearch instance that aggregates structured logs from a fleet of side-project services, letting coding agents triage production errors during development sessions. All demo and test data in this repository is synthetic.

License

MIT © 2026 Kris Cernjavic