I build safe-by-default, open-source engineering across the stack — AI agents, identity, and data platforms — so systems can do real work without being handed the keys.
🌐 dockndevai.github.io — everything on one page, with the security model, live stats, and copy-paste installs. ✍️ I write about how it's actually built on Medium and Dev.to.
Six areas, one principle — least privilege, enforced, not documented:
🤖 AI agent tooling (MCP) · 🔐 Identity & Keycloak · 📚 RAG & LLM platforms · 🧠 On-device AI · 🎙️ Voice · 🧪 Kubernetes labs
A family of Model Context Protocol servers that share one governance model: read-only by default, layered access modes, resource allowlists, protected resources, delete gating, typed confirmation for high-impact ops, secret redaction, dry-run, and JSON audit logging. Mostly MIT-licensed and on npm as @dockndevai/mcp-* (plus the Apache-2.0 ossian-mcp).
| Server | What it does | Install | Stars |
|---|---|---|---|
| mcp-kubernetes | Pods, logs, deployments, scale/restart, apply, exec | npx -y @dockndevai/mcp-kubernetes |
 |
| mcp-openshift | OpenShift / Kubernetes — projects, pods, logs, routes, deployments; token or username/password | npx -y @dockndevai/mcp-openshift |
 |
| mcp-kafka | Topics, consumer groups + lag, create/alter/reset | npx -y @dockndevai/mcp-kafka |
 |
| mcp-clickhouse | Schema, queries, SQL-classified read/write/destructive gating | npx -y @dockndevai/mcp-clickhouse |
 |
| mcp-debezium | CDC connector status, config, lifecycle | npx -y @dockndevai/mcp-debezium |
 |
| mcp-oci | Oracle Cloud discovery + Terraform generation | npx -y @dockndevai/mcp-oci |
 |
| mcp-azure | Azure Resource Manager inventory, tags, VM power, lifecycle | npx -y @dockndevai/mcp-azure |
 |
| mcp-azure-devops | Boards, repos, pipelines, projects | npx -y @dockndevai/mcp-azure-devops |
 |
| mcp-github-enterprise | GitHub Enterprise — orgs, repos, teams, rulesets, security alerts & Actions, with a security-posture baseline | npx -y @dockndevai/mcp-github-enterprise |
 |
| mcp-keycloak | Realms, users, clients, roles, groups | npx -y @dockndevai/mcp-keycloak |
 |
| mcp-percona-pg | Percona PostgreSQL + PgBouncer: pooling, tuning, backups/PITR, DR, upgrades | npx -y @dockndevai/mcp-percona-pg |
 |
| mcp-grafana | Dashboards, datasource queries (PromQL/LogQL/SQL), alerts & annotations | npx -y @dockndevai/mcp-grafana |
 |
| ossian-mcp | Ask your own documents with citations + durable agent memory (RAG) | npx -y ossian-mcp |
 |
| mcp-outlook | Outlook mail + OneDrive files (Microsoft Graph) — read, search, draft, send; browser sign-in | npx -y @dockndevai/mcp-outlook |
 |
| mcp-teams | Microsoft Teams (Microsoft Graph) — teams, channels, chats & messages, post/reply/send; browser sign-in | npx -y @dockndevai/mcp-teams |
 |
| mcp-macos | Observe & operate a Mac — files, processes, apps, screenshots, shell, AppleScript, GUI (safe by default) | npx -y @dockndevai/mcp-macos |
 |
| mcp-mac-control | Full Mac control — drive it like a human: shell, AppleScript, files, and GUI; opt-in safe mode | npx -y @dockndevai/mcp-mac-control |
 |
| mcp-android | Control an Android device over ADB — screenshots, taps, typing, app launch, notifications | npx -y @dockndevai/mcp-android |
 |
| mcp-ios | Drive a physical iPhone/iPad over Appium + WebDriverAgent (XCUITest) | npx -y @dockndevai/mcp-ios |
 |
| mcp-cdp | Drive an Electron/Chrome app over the DevTools Protocol — DOM, console, network, gated JS eval. Cross-OS alternative to computer-use for testing apps | npx -y @dockndevai/mcp-cdp |
 |
| mcp-laya | Fast, local, typed decisions for agents (classify/score/yes-no) in 100+ languages — no hallucination | pipx install mcp-laya |
 |
Each works with Claude, Cursor, OpenAI Codex, VS Code, and Windsurf — per-client setup in every repo's docs/CLIENTS.md.
🛡️ laya-guard — an optional, fully-local AI risk guard for the servers above. High-risk commands (shell,
evaluate, apply/delete/exec) are classified allow/confirm/block by a local decision model before they run — defense in depth that only ever tightens, never grants. Wired into every mutating server in the suite.
🧩 mcp-core — the shared governance library every server above is built on: access modes, the guard, resource allowlists, elicitation confirm, symlink-safe path handling, secret redaction and audit — one audited implementation instead of twenty that drift. On npm as
@dockndevai/mcp-core.
Production Keycloak SPIs in Java — the two things teams hit first: consistent branding across realms, and real tenant isolation.
| Project | What it is |
|---|---|
| keycloak-configurable-theme | Keycloak 26.x SPI: one configurable theme for all realms — layout, colours, logo, favicon, console & email branding via JSON, admin console or REST. |
| keycloak-realm-database-router | Keycloak SPI: one database per realm — hard multi-tenant isolation at the data layer, with a Docker Compose stack and load tests. |
| mcp-keycloak | The agent adapter — multi-realm Keycloak admin over MCP, same safe-by-default model as the suite. |
The agent tools sit on top of real platforms I also build and run.
| Project | What it is |
|---|---|
| ossian | Open-book RAG over your own documents — Spring Boot + Spring AI, a React console, Keycloak OAuth2, Postgres/pgvector and Redis. Answers with citations, plus a maintenance side for ingestion, re-indexing and coverage gaps. |
| ossian-kafka-connect | A Kafka Connect sink for Ossian — turns Debezium CDC events into RAG documents with citations, so the knowledge base stays live. |
| spring-llm-gateway | An LLM gateway on Spring Cloud Gateway (WebFlux): virtual keys, token quotas, usage metering and failover for self-hosted OpenAI-compatible backends (vLLM, Ollama). |
Small models that run where the data is — Laya is a tiny typed-decision model that runs locally (browser, Mac, Raspberry Pi) with nothing uploaded.
| Project | What it is |
|---|---|
| laya-models | The Laya decision model, converted to ONNX so it runs in the browser — no server round-trip. |
| mac-attack | A macOS game & real screensaver: the camera turns people into cartoons and Laya directs harmless toy chaos — ~340 ms/decision on an M2 Pro. Body rectangles only, no face recognition, nothing uploaded. |
| pi-meeting-assistant | A voice-controlled robot arm on a Raspberry Pi: local STT + Laya decisions in English, Hinglish and Hindi — no LLM in the loop. Plus a live meeting assistant. |
| otto-assist | A realtime, voice-driven assistant with Claude as the brain: local Whisper STT, streaming TTS with barge-in, subagents, and a web dashboard. (Homebrew tap) |
Runnable reference environments — every claim measured, every failure documented. Each pairs with an MCP server above.
| Lab | What it covers |
|---|---|
| percona-pg-lab | PostgreSQL + PgBouncer on Kubernetes with the Percona Operator — HA, DR, pooling, backups, observability. Pairs with mcp-percona-pg. |
| kafka-lifecycle-lab | Apache Kafka on Kubernetes with Strimzi — provision, upgrade, DR with MirrorMaker 2, read-only agent triage. Pairs with mcp-kafka. |
| clickhouse-lab | ClickHouse on Kubernetes with the Altinity operator — ReplicatedMergeTree, upgrades, online PVC expansion, backups, cross-region DR. Pairs with mcp-clickhouse. |
Build logs & deep dives on Medium and Dev.to — the bugs, the trade-offs, and the governance decisions behind these systems.
⭐ If any of these save you time, a star helps other engineers find them.