OpenCode Blueprint

One command. A complete, opinionated, production-grade opencode environment — 21 MCPs, 10 plugins, 88 skills, a global agent workflow, and Kodus + PR-Agent review loops. Built for teams shipping Apple-clean, award-worthy software.

npx opencode-blueprint

Restart opencode. Create any folder. Start coding.

Why this exists

Setting up an AI coding environment used to mean weeks of trial and error: which MCP servers, which plugins conflict, which skills are actually S-rank, how to wire review loops so nothing sloppy ever merges. This repo is the finished answer — researched, de-duplicated, and verified working together:

• No conflicts. Every orchestrator, memory, and design overlap was researched and resolved to a single best-of-breed pick.
• Only proven pieces. S-rank skills, official vendor integrations, high-traction community plugins.
• A workflow with teeth. Isolate → Build → Prove → Ship, with a Kodus loop gating every commit and PR-Agent gating every merge.

What's inside

MCPs (21) — live access

| Area | Servers | |---|---| | Project & code | notion, github, figma, sentry | | Docs & design | context7, lucide-animated, expo | | Data | supabase, prisma, postgres, mongodb, redis | | Cloud & infra | gcp-bigquery, gcp-compute, gcp-gke, gcp-run, gcp-storage, cloudflare, kubernetes | | Payments & analytics | stripe, google-analytics |

Plugins (10) — runtime superpowers

| Plugin | Job | |---|---| | oh-my-opencode-slim | 7-agent orchestration (the one orchestrator — swarm/micode/OAC intentionally excluded) | | opencode-mem | Automatic long-term memory with vector recall | | @tarquinen/opencode-dcp | Context pruning (pairs with memory, not against it) | | opencode-review | Auto code-review on idle + /review + auto-fix (local install — not on npm) | | opencode-telemetry | Passive cost/token forensics in local SQLite | | opencode-ralph-wiggum | Self-correcting Ralph loops (/ralph-loop) | | opencode-github-release | Semver tags + gh releases | | envsitter-guard | .env reads blocked, safe key inspection instead | | opencode-log-sanitizer | JWT/bcrypt/base64 redacted before the model sees them | | opencode-command-inject | package.json scripts + Makefiles become / commands |

Skills (88) — stack knowledge

| Stack | Skills | |---|---| | Next.js / React | nextjs-best-practices, vercel-react-best-practices, nextjs-nestjs-integration | | NestJS | nestjs-best-practices | | Prisma / Postgres | prisma-database-setup, prisma-client-api, prisma-cli, prisma-postgres, prisma-postgres-setup, prisma-upgrade-v7, prisma-mongodb-upgrade | | Expo mobile | expo-overview, expo-router, expo-upgrade, expo-native-ui, expo-data-fetching, eas-app-stores, eas-workflows, eas-update | | Tauri desktop | tauri-v2 | | Design (all S-rank) | apple-design, animate, animate-expo, animation-vocabulary, emil-design-eng, review-animations, improve-animations, find-animation-opportunities, pick-ui-library, prototype, design-taste-frontend, impeccable, shadcn | | Workflow | new-feature, code-structure, evidence-driven-testing, before-and-after, unslop, perf-debug, excalidraw-diagram | | API design / contract | api-designer, graphql-architect | | Realtime / sockets | websocket-engineer | | JS/TS | javascript-pro, typescript-pro | | E2E / browser testing | playwright-expert | | Security review / hardening | secure-code-guardian, security-reviewer | | Legacy modernization | legacy-modernizer | | Adversarial review | the-fool | | NestJS deep work | nestjs-expert | | Next.js deep work | nextjs-developer | | React UI logic | react-expert | | Postgres tuning | postgres-pro, database-optimizer | | DevOps / pipeline | devops-engineer | | Test strategy / suite | test-master | | Gate inputs (commands own the gate) | architecture-designer → /architect, feature-forge → /scope, spec-miner → /audit, debugging-wizard → /debug, code-reviewer → review, code-documenter → /document, fullstack-guardian | | Later (surface-gated) | kubernetes-specialist, terraform-engineer, microservices-architect, mcp-developer, monitoring-expert |

The workflow (AGENTS.md, installed globally)

1. Isolate — fresh worktree per feature (established projects; skipped for new apps/migrations); /scope first for new builds/migrations, /audit human-only when context stale.
2. Build — /develop (+ code-structure + stack skills, design-system for UI); → /architect spec when a decision is owed; /debug anytime.
3. Prove — runtime evidence (before/after) + /check verify vs spec, /test owns the suite; then the Kodus loop to clean.
4. Ship — /check review, PR with evidence, /document, /unslop prose, PR-Agent + Kodus at zero (/sync human-only, never auto-chained).

Design bar: Apple-clean, invisible motion (transform/opacity, <300ms, springs, reduced-motion honored), full UI states, dark mode from day one.

Quickstart — new machine

npx opencode-blueprint

The installer: checks prerequisites → backs up your existing config → copies opencode.jsonc + AGENTS.md + commands → installs the local plugin's deps → installs all 88 skills globally → installs the Kodus CLI and redis-mcp-server → prints your personal checklist.

Then:

1. gh auth login · kodus auth login (or KODUS_TEAM_KEY).
2. Set env vars (see config/.env.example).
3. Restart opencode — config loads once at startup.
4. Per repo, once: /pr-agent-setup (scaffolds the PR-Agent Action; key stored as GitHub secrets, never in files).

Preview first: npx opencode-blueprint --dry-run.

Commands (global, work in any project)

| Command | Does | |---|---| | /pr-agent-setup | Scaffolds PR-Agent Action + .pr_agent.toml; asks for LLM endpoint, model, key; stores secrets | | /pr-agent <pr> | /review → fix → re-review until clean | | /kodus-review | Kodus review-fix loop until no high/critical findings | | /ralph-loop "<task>" | Ralph iteration loop (from plugin) | | /review | On-demand code review (from plugin) |

On the PR itself you can also comment /review, /improve, or /ask <question> — PR-Agent answers inline.

Kodus: Cloud vs self-hosted

• Cloud (default): CLI → api.kodus.io, models chosen in the dashboard.
• Self-hosted (kodustech/kodus-ai, Docker): point the CLI with KODUS_API_URL, and give the backend any OpenAI-compatible model via API_OPEN_AI_API_KEY + API_OPENAI_FORCE_BASE_URL + API_LLM_PROVIDER_MODEL. Heads-up: it grabs ports 3000/3001/5432/27017 (collide with local dev DBs), containers can't see host localhost, and telemetry heartbeat is on unless disabled.

Publishing your own fork to npx

npx <name> resolves from the npm registry, so the repo must be published as a package (that's what package.json + bin are for):

npm login                    # once per machine
npm publish --access public  # from the repo root, on every release

Requirements: unique name in package.json, a bin entry pointing at an executable script, npm login done. After publishing, anyone can run npx <your-name> — npx downloads the tarball to a temp dir and executes the bin, no install needed. Version with npm version patch|minor|major before each publish. Prefer scoped names (@your-org/blueprint) to avoid squatting collisions — the bin name stays whatever you declare.

Nothing secret is ever published: .gitignore + files hygiene keep credentials out (keys live in GitHub secrets and shell env, see config/.env.example for the list).

Repo layout

| Path | What | |---|---| | AGENTS.md | Global workflow (source of truth) | | config/opencode.jsonc | 21 MCPs · 9 plugins · instructions | | config/commands/ | /pr-agent, /pr-agent-setup, /kodus-review | | config/plugins/opencode-review/ | Local review plugin sources | | config/.env.example | Required env vars (values never committed) | | bin/setup.mjs | The installer |

Contributing

Issues and PRs welcome. Ground rules: no second orchestrators, no duplicate memory systems, S-rank (or official) skills only, every addition named with its conflict analysis. Run the Kodus loop + PR-Agent to zero before pushing.

License

MIT — see LICENSE.

Acknowledgments

Workflow adapted from michaelshimeles/skills. Design craft: Emil Kowalski, Leonxlnx, Impeccable, shadcn. Data: Prisma, Expo. Review: Kodus, PR-Agent.