Rahdar MCP server
2026년 11월 15일 커밋Apache Artemis, Kafka, RabbitMQ 클러스터를 관리하기 위한 통합 운영 콘솔을 제공하며, 컨슈머 랙(consumer lag) 추정치 및 승인 게이트 기능을 포함합니다.
다른 이름 sudoitir/rahdar
- 스타
- 10
- Tools
- —
- 다운로드, 30일
- —
- 최신 버전
- —
- 마지막 커밋
- 2026년 11월 15일
- 라이선스
- Apache-2.0
Rahdar MCP란?
Apache Artemis, Kafka, RabbitMQ 클러스터를 관리하기 위한 통합 운영 콘솔을 제공하며, 컨슈머 랙(consumer lag) 추정치 및 승인 게이트 기능을 포함합니다.
프로젝트 README에서 생성된 요약
Rahdar MCP server 정보
DevOps 엔지니어와 사이트 신뢰성(SRE) 팀이 여러 메시징 클러스터를 개별 터미널과 관리 콘솔을 통해 관리해야 하는 번거로움을 Rahdar가 해결해 드립니다. Apache ActiveMQ Artemis, Apache Kafka, RabbitMQ를 한곳에서 모니터링할 수 있는 단일 운영 환경을 제공합니다. 서로 다른 시스템을 하나의 뷰로 통합함으로써, 사용자는 도구를 전환하지 않고도 전체 인프라의 상태와 백로그를 추적할 수 있습니다.
이 서버는 OpenBroker SDK를 통해 기존 브로커에 연결되는 통합 콘솔 역할을 합니다. 브로커 자체를 시작하거나 관리하지는 않으며, 대신 인터페이스 내에 기존 클러스터를 등록하여 사용합니다. 보안과 안정성을 위해 시스템에는 파괴적인 작업에 대한 승인 게이트가 포함되어 있어, 변경 사항을 확정하려면 다른 사람의 확인이 필요합니다. 또한 오프셋 리셋(offset reset)과 같은 작업에 대해 미리보기를 제공하여 명령이 실행되기 전 어떤 일이 일어날지 정확히 보여줍니다.
설정 시 PostgreSQL 18 데이터베이스와 함께 서비스를 실행해야 하며, 데이터베이스는 로컬 컨테이너 또는 외부 관리형 인스턴스를 사용할 수 있습니다. 시스템은 Kafka 컨슈머 그룹 랙을 추정치로 추적하고 Artemis 큐에 대한 정확한 깊이(depth) 측정값을 제공하여, 사용자가 확인하는 데이터의 신뢰성을 보장합니다.
활용 사례
- 1클러스터 간 메시지 백로그 모니터링개발자는 모든 Artemis 큐와 Kafka 컨슈머 그룹을 백로그가 큰 순서대로 정렬된 단일 테이블에서 확인할 수 있습니다.
- 2안전한 Kafka 오프셋 리셋 수행엔지니어는 오프셋 리셋을 확정하기 전에 모든 파티션에서 얼마나 많은 레코드가 건너뛰어질지 미리 확인할 수 있습니다.
- 3중단된 Kafka 트랜잭션 해결운영자는 중단된 트랜잭션을 식별하고, 이를 중단하기 전에 어떤 파티션이 해제될지 확인할 수 있습니다.
- 4클러스터 토폴로지 및 상태 감사사용자는 Artemis 클러스터의 복제 상태와 고가용성(HA) 역할을 검증할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
자동 검사
2026년 11월 14일에 실행됨전체 보고서 →아직 인증되지 않았습니다. 서버는 80점 이상의 점수, 심각한 결함 없음, 그리고 검사에 충분한 데이터를 갖추어야 합니다. 검사 방식 · 점수를 높이기 위해 수정할 항목
- 소스 저장소 연결 및 접근 가능sudoitir/rahdar is public and reachablegithub.com/sudoitir/rahdar ↗통과
- 패키지가 이 저장소를 소스로 명시검사되지 않음
- 빌드 출처(Sigstore 증명)검사되지 않음
- 오픈 소스 라이선스 (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-11-14github.com/sudoitir/rahdar ↗통과
- OpenSSF ScorecardOpenSSF Scorecard 6.9 / 10; weakest: Maintained 0, Code-Review 0, CII-Best-Practices 0scorecard.dev/viewer/?uri=github.com/sudoitir/rahdar ↗부분 10/15
- 패키지 이름이 인기 패키지와 유사하지 않음검사되지 않음
- 패키지 및 해당 종속성에 중요한 보안 권고 없음검사되지 않음
- 모델에 대한 숨겨진 지침 없음0 tool descriptions and the README read; no override phrases, credential paths or hidden characters통과
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 원격 엔드포인트가 MCP 핸드셰이크에 응답검사되지 않음
README
GitHub에서 보기Rahdar
Rahdar (Persian راهدار, "road-keeper") is rah (route) + dar (keeper): it keeps every route your messages travel open and safe.
One open-source console for your Apache ActiveMQ® Artemis, Apache Kafka® and RabbitMQ clusters.
Every cluster you run in one estate, lag shown as the estimate it is, a preview before anything destructive, and an approval gate that holds people and assistants alike.
An interactive on-call night · about 2 minutes · runs in your browser
Docs · Quickstart · Kafka · RabbitMQ · Flow · SQL Console · MCP · Discussions
[!WARNING] Alpha. Under active development and not yet feature-complete. Published images are pre-stable dev builds (
sudoit1/rahdar:dev; there is no:latestyet). Expect breaking changes.
Every cluster in one estate: Artemis and Kafka side by side, the worst backlog first.
Quickstart
Rahdar and its Postgres, from the published image. Your brokers stay as they are: you register them in the UI, and nothing here starts one.
base=https://raw.githubusercontent.com/sudoitir/rahdar/main/deploy/compose
mkdir -p rahdar/compose rahdar/postgres && cd rahdar/compose # the compose file mounts ../postgres
curl -sO "$base/compose.prod.yaml"
for f in postgresql.conf 10-pg-stat-statements.sql; do curl -s "${base%/compose}/postgres/$f" -o "../postgres/$f"; done
curl -s "$base/.env.example" | sed \
-e "s|^SECRET_KEY=.*|SECRET_KEY=$(openssl rand -base64 32)|" \
-e "s|^DB_PASSWORD=.*|DB_PASSWORD=$(openssl rand -hex 24)|" > .env
docker compose -f compose.prod.yaml --env-file .env up -d
docker compose -f compose.prod.yaml logs rahdar | grep -A4 'Created administrator'
Open http://localhost:8080 and sign in as admin with the password the last line
prints. It is shown once: you choose your own at the first sign-in, and set up two-step verification, which the
admin role requires.
git clone https://github.com/sudoitir/rahdar && cd rahdar
just up # two replicas behind HAProxy, pinned to the latest release (just up-single: one replica)
docker run -p 8080:8080 \
-e RAHDAR_DB_URL=jdbc:postgresql://db:5432/rahdar \
-e RAHDAR_DB_USER=rahdar \
-e RAHDAR_DB_PASSWORD=... \
-e RAHDAR_SECRET_KEY="$(openssl rand -base64 32)" \
sudoit1/rahdar:dev
Rahdar runs on a managed or highly available PostgreSQL 18 too, in a schema of your choosing and as a role that is not a superuser: see Running Rahdar on an external PostgreSQL.
Every variable, what a reverse proxy needs for the SSE stream, and how to recover a lost first login are in the configuration guide.
Hardware requirements
Measured with a load test on a 4 vCPU / 8 GB machine, Postgres on the same host (sizing guide, which also lists what each size carries and what degrades first). Sizes per host:
| Minimum | Recommended | |
|---|---|---|
| Rahdar node (single instance, or each replica of a replicated deployment) | 2 vCPU, 2 GB | 4 vCPU, 4 GB |
| PostgreSQL 18 (single instance, or the shared server of a replicated deployment) | 2 vCPU, 2 GB, SSD | 4 vCPU, 4 GB, SSD |
A replicated deployment runs two or more Rahdar nodes of that size on one Postgres of that size. The minimum is the observed peak (a Rahdar
replica 1.3 GiB and about 1.2 cores, Postgres 0.6 GB and one core) with little headroom, derived from the runs on the
reference machine rather than measured on a smaller one. It carries a few clusters of up to four nodes with around a thousand queues per node, a few dozen browsers and
about a hundred event-stream subscribers; the recommended size keeps interactive reads under a second
at 3,000 queues per node and absorbs a burst of sign-ins. CPU is what runs out first, not memory. Postgres
needs about 0.6 GB of its memory for the shipped configuration (deploy/postgres/postgresql.conf); how much
disk depends on how long you keep history. The server requirements and settings of an existing PostgreSQL are in
Running Rahdar on an external PostgreSQL.
Why
It's 02:51 and two things page you at once: a Kafka consumer group is falling behind on
payments, and ORDERS.DLQ is backing up somewhere on an Artemis pair. Without Rahdar
that is a management console tab per Artemis broker, kafka-consumer-groups.sh and
kafka-transactions.sh in a terminal, and your own arithmetic in between.
Rahdar puts every cluster you run in one estate, whatever the broker. Health is judged on one scale, and every backlog is in one table, sorted worst first: Artemis queues and Kafka groups side by side, each under its own name. A number says how sure it is: an Artemis queue's depth is exact, a Kafka group's lag is an estimate with its per-partition breakdown, and a number Rahdar cannot read says so instead of showing zero. Nothing destructive runs before you have seen what it will do, and where the approval gate is on, a second person approves it.
▶ Play the story: find the hanging transaction behind the lag, abort it, reset a group with a held approval, and clear the Artemis dead-letter queue, all in one estate.
What it does
| Kafka lag is an estimate, and opens to every partition, slowest first. | A consumer group's lag per partition and its members. |
| An offset reset is previewed per partition before it is held to confirm. | A hanging transaction, found and aborted with its reach stated first. |
| Registration checks every broker on the address it advertises. | Artemis topology, with HA roles polled from every node. |
Rahdar talks to brokers through OpenBroker, an open-source broker SDK with a driver per system. The screens every system shares (the estate, subscriptions and destinations, browsing, purge, the audit trail, the API and MCP) work on each, in that system's own words, and an action a system cannot do stays visible, disabled, with the reason.
| System | Supported | Guide |
|---|---|---|
| Apache ActiveMQ® Artemis | 2.33.0 and later | Guide |
| Apache Kafka® | 2.8 and later, KRaft and ZooKeeper mode | Kafka |
| RabbitMQ | 3.13 and later (4.x tested) | RabbitMQ |
Details and what happens outside those ranges: supported broker versions.
Apache Kafka
- Consumer groups — classic, consumer and share groups with their state, members and lag. Lag is an estimate and is marked as one: open it for every partition, slowest first, with the committed offset, the end and the sample's age.
- Reset offsets — to the earliest, the latest, a time or an offset, previewed per partition with what the group would skip or read again. A group with members is refused, naming them; the reset runs against the state it previewed, is confirmed by holding, waits for approval where the gate is on, and is audited with the offsets before and after.
- Topic configuration — each setting with where its value comes from, changed through a review that the broker validates before anything applies.
- ACLs — listed, added and removed, with the effective access of each principal before and after the change.
- Transactions — a hanging transaction that holds read-committed consumers back is flagged, and aborting it is previewed, held and audited.
- Brokers and the quorum — every broker, whether Rahdar reached it on its advertised address, and the KRaft controller quorum.
- On the shared screens — browse a topic by partition and offset or time, without joining a group; purge a group's unread records or truncate a topic with its reach stated, including every group that would lose records it has not read.
RabbitMQ
- Registration — management URLs and optional AMQP endpoints with one account; the check reaches every endpoint on its own and names each node, its alarms and any partition.
- Nodes — version, uptime, listeners and plugins, with a memory or disk alarm flagged because it blocks every publisher.
- Virtual hosts, bindings, policies and permissions — each change previewed first, confirmed by holding for the destructive ones, held for approval where the gate is on, and audited. A policy's preview lists the queues and exchanges it matches.
- Shovels and federation — dynamic shovels and federation links with their state, addresses shown without credentials.
- On the shared screens — classic, quorum and stream queues with their backlog marked as an estimate (a stream has none); browse a stream by offset, without removing a message; publish through an exchange and be told when no queue received it. A resource recreated with the same configuration is not told apart, and the confirmation says so.
Apache Artemis
-
Topology — live/backup pairs and their replication state. The HA role is polled from every node on every cycle and never read from configuration; two live nodes in one pair raise a split-brain alert.
-
Flow — which application sends to which address, how that address routes through diverts, bridges and cluster hops into queues, and who consumes them, at what rate. Faults such as a backlog with no consumer are stated in words, and clients are sampled only while someone is watching.
-
SQL Console — Artemis cannot tell you where order 4471 went. Its only server-side filter is a JMS selector: headers only, never the body, one queue on one node at a time. Rahdar can:
SELECT * FROM "ORDER.*" WHERE body->>'orderId' = '4471' LIMIT 50The dialect is read-only
SELECT, checked against a fixed column catalogue, so a query cannot change anything. Header predicates become a JMS selector and cost nothing; body predicates scan, and the cost line under the editor tells you which before the query runs. A query over the cost ceiling is refused with its estimate rather than quietly truncated. Add a live tail, or turn on complete capture for a queue: a broker-bounded copy that Rahdar drains, so a message consumed between two polls can still be found. -
Cross-node resources — queues, addresses, consumers, sessions, connections and producers in one virtualised table, each row attributed to its node and kept current over SSE.
-
Message operations — browse, send, move, retry, expire, delete and purge. Every mutating call accepts
?dryRun=true, bulk operations are capped on the server, and outcomes are reported per node. -
Request-reply tracing — requests matched to their replies across addresses and nodes, with latency and timeout statistics against the expectations you declare.
-
Broker configuration — declare the address settings, security settings, diverts, bridges and queues a cluster should run, and compose the routing between them on a canvas. Apply them canary-first, with every hazard stated before anything is written, or export them as a
broker.xmlfragment, and see where each node has drifted. On a first run Rahdar offers the cluster's current state as revision 1; nothing is adopted without your say, and a bridge is never adopted because the broker reports only part of one. -
Setup review — each cluster's HA, clustering, durability and message-safety configuration checked against known mistakes: a single replication pair that cannot win a quorum vote, redistribution left at
-1, a connector advertisinglocalhost, no dead-letter address. Every finding carries the evidence per node and thebroker.xmlfix, and can be accepted as a known risk, with a reason, on the record. -
Alert delivery — Slack, Microsoft Teams, PagerDuty (incidents open and resolve with the alert), email and signed webhooks, each tested before it is saved, with a delivery log that says why a notification failed.
-
Data governance — sensitive headers and properties are masked, PII is classified automatically, and redaction follows the viewer's role.
-
Governance — authentication everywhere, roles and permissions scoped global → environment → cluster, API tokens, optional OIDC/SSO, and an audit trail that records every change and its outcome.
-
MCP server — the same capabilities for an AI assistant, under the same grants and the same audit trail. Its eight
kafka_*tools change nothing that was not previewed first, and wait for a person wherever the approval gate is on; sixrabbitmq_*tools read a RabbitMQ cluster. -
Plugins — install a plugin's
.jarfrom the UI: screens, API, assistant tools and data of its own, with everything it can do shown before you confirm, and no restart unless it needs one. Updates show what changes and roll back; a failing plugin never stops Rahdar. Start a plugin from the template.
How it works
Four rules hold everywhere in the product:
- Broker-friendly by construction. Batched reads (on Artemis, one Jolokia POST per node, never one per queue), tiered polling and a per-node rate limiter. Rahdar must never be the reason a broker falls over.
- Safe by default. Every destructive call can dry-run first and states its reach, and purge, delete, truncate and resets are confirmed by pressing and holding a button that names the count.
- State comes from the brokers, not from config. Artemis HA roles are polled from every node, a Kafka cluster's health is formed from its brokers and its quorum, and a RabbitMQ cluster's from its nodes, their alarms and its replicated queues.
- Honest capability gating. An unavailable feature says so and shows the exact
broker.xmlthat enables it. Nothing silently disappears.
Java 25 · Spring Boot 4.1 · PostgreSQL with Liquibase · React 19 + Vite + Mantine 9 · TanStack Router/Query/Table · React Flow · the OpenBroker SDK and its Artemis, Kafka and RabbitMQ drivers · SSE · one container image. Read the architecture and every recorded decision.
Contributing
Questions and ideas go to Discussions,
bugs to issues, and
vulnerabilities to private reporting
(SECURITY.md). To build it, you need JDK 25, Node 24, Docker and
just; a dev container is included.
just dev-up # Postgres + a real Artemis primary/backup pair + a Kafka node + Rahdar, from source
just dev # or: backend :8080 + Vite :5173 together, with live reload
just verify # everything CI runs
ADMIN_PASSWORD=… just demo (the first run also needs NEW_ADMIN_PASSWORD=…, and later ones ADMIN_TOTP_SECRET=…) adds two more live/backup pairs and fills all six nodes
with realistic traffic: diverts, a bridge, cluster hops, a growing backlog with no
consumer, a real dead-letter backlog and one stopped node. It also registers a
three-broker Kafka cluster, events, with groups that keep up and fall behind, a
compacted topic, ACLs and a hanging transaction. The screenshots above are
recorded from that estate; nothing is staged.
Every feature goes through OpenSpec and significant decisions are recorded as
ADRs. Start with CONTRIBUTING.md.
Releases
A merge that changes the application publishes a release, versioned with CalVer
YYYY.MM.PATCH and tagged three ways on Docker Hub: 2026.09.3 (immutable), 2026.09
(that month) and dev (the latest). There is no :latest until the first stable
release. Each release attaches the runnable jar with its .sha256, and its notes are
generated from the commit messages (releases, changelog). The plugin SDK
(npm) is published with the releases that change it; the plugin API and the Java client reach Maven Central
when a maintainer publishes a release there.
Licence
Apache-2.0, the same licence as Artemis and Kafka themselves.
Apache, Apache Kafka, Kafka, Apache ActiveMQ, ActiveMQ and Artemis are trademarks of the Apache Software Foundation; RabbitMQ is a trademark of its owner. Rahdar is not affiliated with or endorsed by them. The names are used only to describe the brokers Rahdar works with.
If this saved you time, a ⭐ helps others find it.
Rahdar MCP 질문
Rahdar가 브로커 설치를 관리하나요?
아니요, Rahdar는 브로커를 시작하거나 관리하지 않습니다. 기존에 사용 중인 브로커를 UI에 등록하여 사용합니다.
Rahdar는 파괴적인 작업을 어떻게 처리하나요?
파괴적인 작업에 대한 미리보기를 제공하며, 사람과 AI 어시스턴트 모두를 위한 승인 게이트를 포함합니다.
Kafka 랙은 어떻게 표시되나요?
Kafka 랙은 파티션별 상세 내역을 포함한 추정치로 표시됩니다.