LogDoc v2

Structured-log-first platform: a single Go binary on top of ClickHouse. Pipeline: gather → pipe → sink → view → understand.

What makes it different:

• Architecture map from logs alone — services and their dependencies appear on a live map without agents or mandatory tracing;
• MCP built in — AI agents investigate your system through the same interface you use;
• v1 protocol compatible — existing LogDoc appenders keep working.

New here? Start with docs/getting-started.md. Coming from LogDoc v1? See docs/migration-v1.md.

Quick start

docker compose -f deploy/docker-compose.yml up -d --build

UI: http://localhost:9001 · Health: GET /healthz

First log:

curl -X POST localhost:9001/api/v1/ingest \
  -d '[{"msg":"hello logdoc","app":"demo","lvl":"INFO","fields":{"env":"dev"}}]'

Search:

curl 'localhost:9001/api/v1/query?app=demo&lvl=INFO&q=hello'

Live tail: WebSocket GET /api/v1/tail (same filter parameters).

v1 compatibility: TCP/UDP :9999 accepts the ld_format protocol — existing logdoc-go-appender and logback-appenders work unchanged.

OTLP: gRPC logs on :4317 — point any OpenTelemetry SDK or Collector at it (service.name→app, body→msg, severity→lvl, attributes→fields).

Syslog: RFC 3164 and RFC 5424 with auto-detection, TCP (newline or octet-counting framing) and UDP. Off by default — enable with LOGDOC_SYSLOG_UDP_ADDR=:5140 / LOGDOC_SYSLOG_TCP_ADDR=:5140 (or the ingest.syslog section of the config) and point rsyslog, a router or a NAS at it.

Journald: the systemd journal export format over UDP (LOGDOC_JOURNALD_UDP_ADDR=:5514), including binary fields:

journalctl -o export -f | socat - udp-sendto:logdoc-host:5514

Python: the stdlib logging handlers speak to LogDoc directly (LOGDOC_PYTHON_TCP_ADDR=:9020) — no library, no formatter:

import logging, logging.handlers
logging.getLogger().addHandler(logging.handlers.SocketHandler("logdoc-host", 9020))

Architecture map

LogDoc builds a live service map from the logs themselves: shared trace_id/correlation_id values and peer fields (peer.service, target, upstream) turn into directed edges between services. No agents, no tracing required — traces only refine the map.

• UI: the Topology tab — force-directed map, click a service or an edge to see its details and jump to its logs.
• API: GET /api/v1/topology?window=5m — nodes and edges with windowed rates (rps, error rate).
• Export: GET /api/v1/topology/export?format=mermaid|markdown — paste the current architecture straight into your docs.
• Deploy markers: version changes detected from the logs themselves (a version field, or a deploy message) land on the service card next to whatever happened right after — "2.3.1 deployed, seconds later the first errors". API: GET /api/v1/deploys?app=billing&window=24h.
• What changed: GET /api/v1/topology/diff?window=1h — new and silent services and links, error-rate jumps versus the previous window, and deploys, in one report. In the UI: the Changes toggle on the map.

MCP: the agent interface

LogDoc is an MCP server: any agent (Claude Code, or anything speaking MCP over Streamable HTTP) can investigate your system through four tools — query_logs, get_topology, get_topology_diff, get_service_card.

claude mcp add --transport http logdoc http://localhost:9001/mcp \
  --header "X-API-Key: <your key>"

Then ask the agent things like "why is checkout failing?" — it walks the map, follows the error edges and reads the logs itself. Try it on the demo incident: deploy/demo-incident.sh injects a database failure cascading through three services.

Above (2× speed): a single prompt, no human input — the agent walks the topology, follows one trace_id across four services and lands on the root cause: a bad billing deploy exhausting the postgres connection pool.

Pipelines: structure on ingest

Server-side pipelines parse raw messages into structured fields the moment they arrive — before storage, search, live tail, the topology extractor and alert rules see the entry. A raw nginx access line becomes queryable fields, gets its real severity, and puts the upstream service on the map:

pipelines:
  - name: nginx access
    when: { app: nginx }
    steps:
      - grok: '%{COMBINEDAPACHELOG} upstream=%{WORD:upstream}'
      - severity:
          from: response
          rules:
            - { prefix: "5", lvl: ERROR }
            - { prefix: "4", lvl: WARN }
            - { lvl: INFO }
      - set:
          fields: { peer.service: $upstream }   # nginx→upstream edge on the map
  - name: drop health checks
    when: { msg_regex: 'GET /healthz ' }
    steps: [{ drop: true }]

Steps: grok (built-in pattern library, no external dependency), regex (RE2 named groups), json (JSON messages → dot-flattened fields, honoring message/level keys), severity (level from a field value or from level names like warning/fatal), set (rewrite app/src/pid/msg/lvl or add fields, $name references), drop. A when selector limits a pipeline by app, src or msg_regex. See logdoc.example.yml.

Notifications

Built-in alert rules run over the live stream — no query polling:

• error_threshold — N entries with level ≥ ERROR within a sliding window;
• silence — a service that used to log stopped logging.

An error_threshold rule can carry a composite match condition instead of the default "level ≥ ERROR": nested and/or, exact app/src/pid, minimum lvl, string ops on the message (contains/starts/ends/equals, case-insensitive by default), an RE2 regex, and the same ops on parsed key=value fields. max_fires retires a rule after N alerts (0 = fire once). Alerts include the matched log entries.

Events go to Telegram, a webhook (JSON POST — the integration point for everything else), email and/or a Kafka topic. Configure in logdoc.yml:

notify:
  rules:
    - name: web went silent
      type: silence
      app: web
      window: 5m
    - name: billing cascade
      type: error_threshold
      threshold: 2
      window: 1m
      match:
        app: billing
        or:
          - lvl: ERROR
          - msg: { contains: "pool exhausted" }
  telegram: { token: "...", chat_ids: [123456789] }
  webhook:  { url: "https://example.com/hook" }
  kafka:    { brokers: ["localhost:9092"], topic: "logdoc-alerts" }

Rules also live in the UI: the Rules tab lists every rule with live fire counters and lets you create, edit and delete rules at runtime (persisted server-side; config-file rules are read-only there). The same works over GET/POST/DELETE /api/v1/notify/rules.

See logdoc.example.yml for every option.

Users and roles

With no users and no API key everything is open — good for a laptop. For a team, set ingest.api_key (the bootstrap admin credential), sign in with it and create accounts on the Access tab (or POST /api/v1/users):

• member — search, tail, topology, export;
• admin — everything, plus notification rules and user management.

Users sign in with login+password (JWT session) and issue personal ldt_... tokens for scripts and CI — revocable per token, carrying the user's role. Every credential is sent the same way: X-API-Key, Authorization: Bearer, or ?api_key=. Creating the first user turns auth on everywhere, including OTLP ingest; the API key remains the recovery path.

Plugin SDK

External plugins are standalone executables the core launches and supervises over gRPC — a crashed plugin is restarted with backoff, and a plugin can be written in any language that speaks gRPC. Two kinds:

• source — receives data by its own means (a socket, a queue, an API) and streams entries into the core;
• pipe — becomes a notification channel, usable in alert rule channels next to telegram/webhook/email/kafka.

plugins:
  - name: syslog-plugin
    kind: source
    exec: ./bin/plugins/syslog-source
    config: {udp_addr: ":6514"}

The contract is one proto file — pkg/sdk/proto/plugin.proto; Go plugins use the pkg/sdk helpers (sdk.ServeSource / sdk.ServePipe) and are a single small main.go. Reference implementation: plugins/syslog-source, built with make plugins.

Development

make up      # ClickHouse for development (ports 8124/9010)
make ui      # build the frontend (ui/dist, goes into go:embed)
make build   # bin/logdoc
make test
./bin/logdoc # config: flags / env LOGDOC_* / -config logdoc.yml (see logdoc.example.yml)

Auth: LOGDOC_API_KEY (X-API-Key header, Authorization: Bearer, or ?api_key=). An empty key and no users means dev mode with no auth (see "Users and roles").

License

Apache 2.0