MQ-Sentinel

io.github.pramodreddyboddu由 GitHub 账号 pramodreddyboddu 发布v0.5.0更新于 Oct 24, 2026

Read-only IBM MQ diagnostics for AI agents. Cited root cause, never destructive.

概览

AI 生成的概览

只读的 IBM MQ 诊断 MCP 服务器,让助手运行八项排障检查并获得带引用的根因与修复步骤。

可以做什么
  • 诊断通道故障,包括授权与连接错误、INDOUBT 状态及 AMQ 错误码
  • 按原因码分析死信队列并识别回退循环,仅读取消息头
  • 检查集群健康、过期的集群成员与异常的集群通道
  • 诊断 Native HA、RDQM 与多实例队列管理器的副本延迟、仲裁与脑裂问题
  • 检查 z/OS 队列共享组的 CHIN、页集与缓冲池
  • 执行综合健康检查,返回排序后的总体摘要
适用场景
适合运行 IBM MQ 的团队,希望让助手在不授予写权限的前提下排查通道故障、死信队列、集群、高可用或 z/OS 问题。主要面向需要可引用、可复现诊断结果的企业 MQ 运维与平台团队。
运行要求
以 stdio 在本地运行,通常通过 uvx 从 PyPI 包启动。默认使用内置演示队列管理器;连接真实队列管理器需要清单目录、密钥目录,以及带 IBM MQ 客户端库的 pymqi 附加组件。生产环境可使用带 OIDC 的 HTTP 传输。环境变量包括 MQS_SERVER_ENVIRONMENT。
安装前请注意
它会用密钥目录中的凭据连接真实队列管理器,应将该账号限制为仅显示权限。本地开发开关 MQS_AUTH_DISABLE_AUTH_FOR_LOCAL_DEV 会关闭认证,仅在环境为 prod 时被拒绝。README 中还有把远程脚本直接管道给 shell 的安装命令,执行前请先审阅。

安装

在 SourceWeft 中

  1. 打开 控制台中的 MQ-Sentinel,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

MQ-Sentinel

Production-grade, read-only IBM MQ diagnostic MCP server that AI agents can safely use — with zero hallucinations.

I built this as a serious side project to solve a real, high-stakes problem in enterprise environments. It is now a strong public portfolio piece demonstrating systems engineering, security, observability, production deployment, and building safe infrastructure for AI agents.

[Python 3.12+] [mypy strict] [ruff] [tests 271 passing] [Org-Ready]

0.5.0 — 8 diagnostic tools across all 10 IBM MQ flavors. Read-only by design. Prompt-injection firewall. OIDC + RBAC. Hash-chained audit. Verified IBM Knowledge Center citations (CI-enforced). Production Helm with HPA + air-gapped packaging.

GitHub: https://github.com/pramodreddyboddu/mq-sentinel
Live Demo: https://mq-sentinel.io
Get it in front of people: docs/LAUNCH.md


The Problem I Solved

Enterprise IBM MQ teams waste hours diagnosing issues like:

  • 2035 NOT_AUTHORIZED storms
  • Exploding DLQs
  • Native HA replica lag and split-brain
  • Stale cluster members
  • z/OS mysteries

Traditional tools require deep expertise and manual log diving. Handing an LLM raw MQ access is unacceptable for security and compliance teams.

MQ-Sentinel gives Claude, Cursor, Grok, and other agents a safe MCP interface that returns accurate, cited diagnostics without ever hallucinating or executing destructive commands.


The Problem

IBM MQ teams (banks, telcos, insurance, gov) waste hours on:

  • 2035 NOT_AUTHORIZED channel storms
  • Exploding DLQs
  • Native HA replica lag / split-brain
  • Stale cluster members
  • z/OS QSG mysteries

Traditional tools require deep MQSC knowledge and grepping logs. Giving an LLM direct MQ access is a non-starter for security/compliance teams.

MQ-Sentinel solves this by giving Claude, Cursor, Grok, etc. a safe, read-only MCP interface that returns typed Root Cause + Fix Steps + verified IBM docs — never fabricates, never executes destructive commands.


🎬 See it in 90 seconds

bash
git clone https://github.com/pramodreddyboddu/mq-sentinel.gitcd mq-sentinelmake demo

Watch MQ-Sentinel diagnose a real 2035 NOT_AUTHORIZED, an INDOUBT channel, a 1247-message DLQ, and a Native HA replica disconnect — all against the bundled fixture sandbox. No live IBM MQ required.

[asciicast]   [Try live demo]

Want to record it for sharing? make demo-record produces a browser-playable asciinema cast.

Quick self-check (no MQ needed):

bash
uv run mq-sentinel doctoruv run mq-sentinel tools     # list all diagnosticsuv run mq-sentinel info      # overview + security highlights

Use it from Claude, Grok, Cursor, Gemini, ChatGPT

One MCP server — not a per-vendor plugin. Copy-paste configs: docs/mcp-clients.md.

List it where agents discover servers: official MCP Registry (server.json) + Smithery (smithery.yaml). ChatGPT needs a hosted HTTPS endpoint, not local stdio.

Architecture at a Glance

mermaid
flowchart TB    subgraph Client["MCP Client (Claude / Cursor / Grok)"]        P[Prompt]    end
    subgraph Sentinel["MQ-Sentinel (read-only)"]        OIDC[OIDC + RBAC]        RL[Rate Limiter]        DISP[Tool Dispatcher]        ALLOW[MQSC Allowlist]        CONN[Connector<br/>(pymqi / fixture)]        RCS[RCS Engine +<br/>KC Registry]        SAN[Output Sanitizer]        AUDIT[Hash-chained<br/>Audit Log]    end
    subgraph MQ["IBM MQ (read-only service account)"]        QM[Queue Managers]    end
    P -->|MCP stdio / HTTP| OIDC    OIDC --> RL --> DISP    DISP --> ALLOW    ALLOW --> CONN    CONN -->|raw output| RCS    RCS --> SAN    SAN --> AUDIT    CONN --> QM    QM -.->|DISPLAY only| CONN

Safe by design. No LLM inside the server. Every citation verified in CI.


Why This Stands Out as a Portfolio Project

This is not a toy or weekend demo. It demonstrates real engineering depth:

  • Systems + Security Engineering: Full threat model, multi-layer read-only enforcement, prompt-injection firewall, OIDC+RBAC, tamper-evident audit, hardened containers, SBOM + signed images.
  • Production Operations: Production Helm (HPA, ServiceMonitor, NetworkPolicy), air-gapped packaging (RPM/DEB), observability out of the box.
  • AI Agent Infrastructure: A complete, safe MCP server (stdio + HTTP) designed for real use by agents like Claude, Cursor, and Grok.
  • Enterprise Domain Expertise: Deep coverage of IBM MQ across 10 topologies with version-aware, verified citations.
  • Systematic Delivery: Executed a full 7-phase org-readiness plan with compliance artifacts that actual organizations need.

Built with Grok Build (primary coding partner) while maintaining strict engineering standards.

Full journey and plan: docs/ORG-READINESS-PLAN.md

See also: VISION.md · CONTRIBUTING.md · DEVELOPMENT.md


The Journey (Portfolio Story)

I built MQ-Sentinel to solve a painful, recurring problem in enterprise IBM MQ environments — and I treated the entire effort as a serious engineering and portfolio project.

I used Grok Build (xAI) as my primary coding partner from architecture through implementation, docs, Helm, compliance artifacts, and polish. Every major decision was deliberate: read-only first, citations that can't go stale, defense-in-depth, production packaging.

Key milestones:

  • Designed threat model + multi-layer security architecture before writing the first tool
  • Delivered 8 diagnostic tools across 10 MQ flavors with CI-enforced IBM Knowledge Center citations
  • Production deployment (Helm + HPA + ServiceMonitor, distroless, air-gapped RPM/DEB)
  • Full 7-phase org-readiness plan executed (see links above)
  • 271+ tests (including security-negative tests), strict mypy, comprehensive docs for platform teams

This shows end-to-end ownership: deep domain knowledge, security by design, production operations, and shipping artifacts that real organizations can actually use.

This is my public portfolio project. It demonstrates systems engineering, security, observability, AI infrastructure, and disciplined delivery.

Built with Grok Build while holding high standards throughout.

Security posture — baked in, not bolted on

  • Read-only always. Static MQSC allowlist (DISPLAY / DIS / PING CHANNEL only). Destructive verbs are rejected by three layers: tool, connector, MQ-side setmqaut.
  • Prompt-injection firewall. Every MQ-sourced string is sanitized (control chars, zero-width, ANSI, unicode tag chars, jailbreak markers) and wrapped in a quarantine envelope before leaving the server.
  • URL allowlist + live verification. Responses may only cite www.ibm.com; all other URLs are redacted. Every citation in the registry (20+ reason codes, 8 AMQ codes, 18 topic pages) is fetched daily in CI — a dead link fails the build, not the customer.
  • Tamper-evident audit log. Hash-chained JSONL; mq-sentinel verify-audit detects any retroactive edit.
  • OIDC + RBAC. Principals scoped to nonprod-read / prod-read / admin-audit. Cannot query prod QMs from a nonprod token.
  • Hardened runtime. Distroless image, non-root, read-only FS, dropped capabilities, seccomp, network egress allowlist.
  • Supply chain. SBOM, cosign-signed images, pip-audit + Trivy in CI, every PR security-gated.

See SECURITY.md and docs/threat-model.md.

Org / Enterprise Ready?
This project was deliberately built for real organizations (banks, gov, large enterprises). See the complete journey:

See VISION.md for the owner product principles and long-term thinking.


Skills & Engineering Rigor (Portfolio Highlights)

What this project demonstrates:

  • Systems & Security Engineering — Full threat model, prompt-injection defense, multi-layer read-only enforcement, OIDC+RBAC, hash-chained audit, distroless hardening.
  • Observability & Production Operations — Metrics, dashboards, alerts, runbooks, air-gapped packaging, Helm productionization.
  • AI Agent Infrastructure — Built a real, safe MCP server (stdio + HTTP) that LLMs can use without risk.
  • Enterprise Systems — Deep IBM MQ knowledge across 10 topologies + version-aware diagnostics.
  • Software Craftsmanship — 271+ tests (including security), strict mypy, ruff, CI-driven doc verification, SBOM + cosign signing.
  • End-to-End Delivery — From idea to org-ready artifacts, including comprehensive documentation that platform teams can actually use.

Built iteratively using Grok Build as the primary coding partner while maintaining high engineering standards.

Supported IBM MQ flavors (Phase 1 complete)

Standalone · Multi-Instance QM · RDQM · Native HA · Native HA + CRR · Uniform Cluster · Traditional Cluster · z/OS Queue Sharing Group · MQ Appliance · Containerized.

Diagnostic tools

ToolCoversDemo finding
diagnose_failed_channelsDistributed channels2035 NOT_AUTHORIZED, 2009/2059 connection errors, INDOUBT, AMQ9202/9208/9503
analyze_dlq_and_suggest_reprocessingDLQ (headers only — never bodies)Grouped by reason 2035/2080/2030/2051/2053/2079, backout-loop detection
check_cluster_healthTraditional + uniform clusterPartial repository, stale CLUSQMGR, suspended members, unhealthy cluster channels
diagnose_native_ha_issuesK8s/OpenShift Native HAReplica state, quorum, log replay lag, split-brain, CRR lag
diagnose_rdqm_issuesOn-prem RHEL RDQMPacemaker quorum, offline nodes, DRBD connection/disk, split-brain
diagnose_zos_qsg_issuesz/OS Queue Sharing GroupQSG members, CHIN, page sets, buffer pools, CF structures
diagnose_multi_instance_issuesTraditional MIQMActive/standby state, dual-active split, standby permission, failover events
full_mq_health_checkAll of the above (composite)Executive summary + ranked findings

Supported MQ versions

9.2 LTS, 9.3, 9.4 (incl. 9.4.4+), z/OS. Version is auto-detected at connect time; KC doc links are keyed to the detected version.

Install — pick your path

ScenarioCommandTime
Claude Code plugin/plugin marketplace add pramodreddyboddu/mq-sentinel then /plugin install mq-sentinel@mq-sentinel (details)1 min
Solo / startup laptop (dev)curl -fsSL https://raw.githubusercontent.com/pramodreddyboddu/mq-sentinel/main/scripts/install.sh | MQS_DEV_MODE=true MQS_DEV_MODE_ACK_INSECURE=yes bash5 min
Solo / startup (prod)Same, with MQS_AUTH_OIDC_* env vars exported5 min
Mid-org Kuberneteshelm install mq-sentinel oci://ghcr.io/pramodreddyboddu/charts/mq-sentinel --set oidc.issuer=… --set oidc.audience=… --set oidc.jwksUrl=…30 min
Local K8s POC (no live MQ)cd examples/kind && ./install.sh5 min
RHEL / Rocky / OELsudo dnf install https://github.com/pramodreddyboddu/mq-sentinel/releases/latest/download/mq-sentinel-0.4.0-1.x86_64.rpm5 min
Debian / Ubuntusudo apt install ./mq-sentinel_0.4.0_amd64.deb5 min
Air-gapped (banks/gov)Mirror RPM internally, sign with org GPG, deploy via Satellite / Aptly1 evening

Full guide: docs/INSTALL.md. For IBM MQ client libs: docs/byom.md.

Connect real queue managers

Out of the box MQ-Sentinel serves the bundled demo QM (DEMO_QM). To diagnose your own QMs, set MQS_SERVER_INVENTORY_DIR (QM list) and MQS_SERVER_SECRETS_DIR (credentials), and install the mq extra (pymqi, which needs the IBM MQ client). The server then connects live over a display-only channel. docs/onboard-new-qm.md has the least-privilege setmqaut / CHLAUTH setup.

Quick start (developer / contributor)

bash
make install        # uv sync + editable installmake test           # pytestmake ci             # lint + type + tests + security suite (everything CI runs)make docker         # build the production container imagemake rpm deb        # build RPM + DEB packages (requires `gem install fpm`)

See:

Transports

bash
# stdio (default — for Claude Desktop, Cursor, Claude Code)uv run mq-sentinel serve
# HTTP (production — OIDC bearer auth, /healthz, /readyz, /metrics)uv run mq-sentinel serve --transport http --host 0.0.0.0 --port 8080

See docs/http-transport.md for the OIDC config, endpoint reference, and operational notes.

Repository layout

src/mq_sentinel/  security/     command allowlist, output sanitizer, rate limiter  audit/        hash-chained JSONL audit logger  auth/         OIDC + RBAC  inventory/    QM registry (pluggable)  secrets/      pluggable secrets backend (default: filesystem/K8s)  connectors/   MQ connection layer (pymqi + fixture)  rcs/          Root Cause Summary engine + KC doc registry  tools/        MCP tools (populated Phase 1+)  telemetry/    OpenTelemetry + structured logging  cli/          control-plane CLIdeploy/  Dockerfile    distroless, non-root, read-only FS  helm/         Kubernetes chart (security defaults on)tests/  security/     allowlist + injection firewall corpus (non-negotiable)  unit/

Org-Readiness (7-Phase Plan Executed)

This project was deliberately taken through a complete org-readiness program (7 phases) so it is credible for real platform/SRE teams. See the full plan and status:

Current state: Ready for platform team evaluation.

Technical Roadmap

Phase 1 (MCP diagnostic tools) is complete. See CHANGELOG.md for future technical items:

  • 0.4.0 — ServiceNow / Jira ticket auto-draft from RCS findings.
  • 0.5.0 — SOC 2 evidence-pack generator.
  • Historical telemetry, air-gapped improvements, etc.

Safe remediation is intentionally out of scope. MQ-Sentinel only reports.

For Organizations

MQ-Sentinel is designed from the ground up for enterprise use:

  • Read-only, prompt-injection resistant, OIDC + RBAC
  • Verifiable IBM Knowledge Center citations
  • Production deployment options (Kubernetes, RPM/DEB, air-gapped)
  • Full audit trail

Start here:

Helm chart includes production values, HPA, monitoring, and network policies.

License

Proprietary — © 2026 MG. See LICENSE. IBM® and IBM MQ® are trademarks of IBM Corporation; MG is not affiliated with IBM.

来源:README.md,提交 7cb2b2f

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

3
  1. v0.5.0最新Oct 9, 2026
  2. v0.4.0Oct 9, 2026
  3. v0.3.0Sep 16, 2026

评分

已满足适用检查项的 40%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分

    请在仓库或注册中心条目中声明许可证。