Azure DA.live MCP — Monorepo

A personal monorepo of AI-powered content authoring, migration, and evaluation tools for da.live (Adobe Edge Delivery Services), built on Claude and the Model Context Protocol (MCP).

Start here: REFERENCES.md (repos, working demo, docs, tools used) · docs/architecture/ (the system diagram by diagram, including how the agentic parts work)

The repository's center of gravity is v2.0 — the agents/ A2A Agent Platform: a decoupled mesh of independently-addressable AI agents that generate, migrate, and evaluate content. The other projects are the supporting cast — the MCP server it authors through, the prompt library, the API/admin tooling, and the original evaluation app (now a frozen v1.x backup).

Two version lines. v2.0 = the new agents/ platform (flagship, deployed on Cloudflare, the only actively developed workstream). v1.1.0 = the legacy content-authoring-eval Next.js app on Oracle — frozen and untouched as the safety net (decision D5). Its eval engine was copied into agents/eval-service, not moved.

⚠️ Deprecated subprojects (2026-06-27). content-authoring-eval/ and agent-claude-sdk/ are officially deprecated — superseded by the agents/ v2.0 platform. content-authoring-eval/ stays running on Oracle as a frozen backup (do not modify — D5); agent-claude-sdk/ was prototyping scaffolding whose patterns now live in the platform. Neither receives new work. All new development happens in agents/.

---

⭐ The flagship: agents/ — A2A Agent Platform (v2.0)

Status: v2.8.1, deployed on Cloudflare Workers + Containers (see CHANGELOG.md); the closed loop runs end-to-end with Kimi (K3 default) or Claude (Agent SDK) really authoring da.live pages, scored by the real agentic eval in three modes (fidelity / quality / redesign); the coordinator ships its own Next.js dashboard (Google SSO) on :4004 / content-factor-dash.jackzhaojin.com. An agent-led daily content loop (GitHub Actions cron) generated a fresh Wilderness Journal article each day onto the adapt-to-2026-demo site through the adaptTo() Sept-2026 demo (paused since 2026-10-24: the workflow is disabled on GitHub, not deleted), and a model-matrix benchmark runs the same migration across N models (2026-08-29 result: K3 88 > Opus 5 86 > K2.7 85 > Sonnet 5 / Haiku 4.5 84). Purpose: A multi-agent mesh speaking the A2A protocol (official @a2a-js/sdk), where each agent is its own Express server with an Agent Card, Task lifecycle, and streaming.

The headline capability is a closed loop: the coordinator asks content-gen to fabricate a synthetic "legacy" page, hands it to the migration agent to author into da.live, then to the eval agent to score the result across four dimensions — fanned out and aggregated into variance stats. But the coordinator routes intelligently: it can also just evaluate, just migrate, generate+migrate, or auto-decide; it need not start at generation or end at evaluation.

Agent	Port	Does
coordinator	4004	Routes, fans out, aggregates variance (A2A client and server) — plus its own Next.js dashboard at :4004/ (trigger, live activity feed, branch grid)
content-gen	4002	Content briefs + synthetic legacy source pages (template tier)
migration	4003	Authors into da.live — one Agent Card over backends (dryrun / sdk = Claude via Agent SDK OAuth, the default since 2026-10 / opencode = Kimi K3, optional - no Kimi subscription backs it any more / Make.com), all real-page verified
eval	4001	4-dimension migration-quality evaluation (engine copied from v1.x; deterministic always, agentic when Claude creds are set; modes fidelity / quality / redesign)

Quick Start:

cd agents
nvm use 20 && npm install
cp .env.example .env && set -a; source .env; set +a   # secrets gitignored
npm run dev:eval & npm run dev:content-gen & npm run dev:migration & npm run dev:coordinator &
npm run loop -- "rooftop solar maintenance" --fan-out 2   # drive the closed loop
npm run loop -- "Chasing light on an alpine lake circuit" --backend opencode --site adapt-to-2026-demo --owner jackzhaojin  # Kimi authors a REAL Wilderness Journal article
npm run model-matrix                                      # benchmark N models over the same migration + eval
# coordinator dashboard: http://localhost:4004/

Documentation:

• agents/README.md — overview + status · agents/CLAUDE.md — dev hub
• Build report: ai-docs/2026-06-08-a2a-platform-v2.0/ — as-built architecture + sequence diagrams
• Plan: ai-docs/2026-06-05-a2a-agent-platform/ — the PRD (decisions D1–D6)

Key Features:

• Official A2A SDK end-to-end: Agent Cards, message/stream (SSE), tasks/get, push notifications, an edge webhook shim
• Model-vendor-swappable migration: the same contract, MCP server, and skill run under dryrun, Make.com, Claude via the Agent SDK (subscription OAuth, per-run model; the default since 2026-10), or Kimi headless via opencode (K3, per-run model selection; kept for anyone with a Kimi subscription, none backs this project any more) — real da.live pages authored, published, and scored, and benchmarked head-to-head by npm run model-matrix (docs)
• Coordinator dashboard (Next.js 15 riding the same :4004 process — A2A wire surface untouched): trigger runs (single / bulk / direct-eval / migrate-a-real-page lanes, per-run model + guiding principles), watch live tool/skill activity (sdk/claude-opus-5 → mcp__dalive__save_dalive_content), branch grids, variance tables, JSON export
• Persistence on Cloudflare D1 (same SQL as local SQLite) + artifacts on R2 (public r2.dev)
• Make.com interop remains available in the code; the legacy local tunnel (a2a.xpri.ai) was retired on 2026-10-26. A new public callback route is needed to use that local workflow again.
• Browser-pooled, job-queued, restart-survivable eval; deterministic always + agentic when Claude creds are configured
• Three real-server test tiers (fast/CI, live, soak) — no mocks

---

Supporting projects

functions/ — Azure Functions MCP Server

Status: Production-ready · Node 22+ HTTP MCP server for AI-assisted content editing on da.live — the server the migration agent authors through (decision D1: kept over Adobe's da-mcp). Six MCP tools (list/get/save/create/folder/preview-publish), multi-LLM, Claude Desktop stdio bridge.

cd functions && npm install && npm start

functions/README.md · functions/CLAUDE.md

content-authoring-eval/ — CMS Migration Evaluator (v1.x, DEPRECATED · FROZEN)

Status: ⚠️ Deprecated (2026-06-27) — superseded by agents/eval-service (its engine was copied out). Still running on Oracle Cloud as the frozen safety net; do not modify (D5), and its deploy workflow must never be triggered by platform work. content-authoring-eval/README.md

agent-claude-sdk/ — Agent SDK Experiments (DEPRECATED)

Status: ⚠️ Deprecated (2026-06-27) — prototyping scaffolding (CLI chat, PDF generator, an earlier eval prototype, third-party demos) that was the blueprint for several platform pieces. Those patterns now live in agents/; this folder receives no new work. agent-claude-sdk/README.md

make-dot-com/ — Make.com Agent Prompts

Status: Active versioning Progressive prompt files for the EDS migration agent on Make.com (MVP → +Memory → +BlockLibrary → Full). Copy-pasted into Make.com's UI, not deployed via Git. The platform's migration agent can call these Make.com scenarios via its makecom backend (the default backend is now opencode/Kimi). make-dot-com/README.md

bruno/ — API Testing Collections

Status: Active use Bruno HTTP collections for da.live Admin API and Azure Functions endpoints. bruno/README.md

hlx-admin/ — AEM Admin API Execution Logs

Status: Active use Auditable, one-at-a-time AEM Edge Delivery admin operations — dated working dirs with EXECUTION.md plans + request/response artifacts, driven by the hlx-admin-api-executor skill (GET/SET/GET + human approval).

references/ — POCs & spikes

De-risking spikes referenced by the platform: references/cloudflare/ (long-SSE-through-Containers, container→D1 worker-proxy), references/kimi/ (Kimi K2.6 / opencode backend), references/claude/ (Agent SDK POCs).

---

Repository Structure

azure-da-mcp/
├── agents/                    # ⭐ v2.0 A2A Agent Platform (Node 20, npm workspaces)
│   ├── a2a-common/            #    shared bootstrap (server, stores, client, migrations)
│   ├── eval-service/          #    :4001 eval agent (engine copied from v1.x)
│   ├── content-gen/           #    :4002 briefs + synthetic sources
│   ├── migration-agent/       #    :4003 swappable backends (dryrun/opencode·Kimi/sdk·Claude/makecom)
│   ├── coordinator/           #    :4004 routing + fan-out + variance + CLI + Next.js dashboard (the sole UI)
│   ├── deploy/                #    Cloudflare Worker + 4 Dockerfiles (standalone, Node 22)
│   ├── contracts/             #    JSON Schemas for every skill
│   ├── store-mcp/             #    stdio MCP — conversational store queries
│   ├── e2e/                   #    real-server tests (fast/live/soak/cloud) + the model-matrix harness
│   └── docs/                  #    r2-setup · tunnel-setup · makecom checklist · model-matrix
├── functions/                 # Azure Functions MCP Server (Node 22)
├── content-authoring-eval/    # v1.x eval app — FROZEN backup (Node 20, Docker)
├── agent-claude-sdk/          # Agent SDK experiments (TypeScript)
├── make-dot-com/              # Make.com agent prompts (Markdown)
├── bruno/                     # API testing collections (Bruno)
├── hlx-admin/                 # AEM admin API execution logs (skill-driven)
├── references/                # POCs & spikes (cloudflare, kimi, claude)
├── ai-docs/                   # Planning PRDs + as-built reports (public, indexed in ai-docs/README.md)
├── CHANGELOG.md               # Per-minor-version history of both release lines
├── RELEASES.md                # Release strategy and versioning
└── README.md                  # This file

Releases & Versioning

Lockstep SemVer, trunk-based releases (tag directly from main) — strategy in RELEASES.md, per-version history in CHANGELOG.md.

• v1.1.0 — the legacy content-authoring-eval app (deprecated, frozen backup). v1.* tag push triggers the Oracle deploy.
• v2.x — the agents/ platform (the active line, currently v2.8.1). A major bump because it's a ground-up re-architecture. Deployed to Cloudflare Workers + Containers via .github/workflows/deploy-agents.yml on v2.x+ tags (version tracked in agents/package.json).

Common Dependencies

• Node.js: 20.x (agents/, content-authoring-eval/), 22.x (functions/; also required for the Wrangler/Cloudflare CLI)
• A2A SDK: @a2a-js/sdk (the platform's protocol layer)
• Claude / Anthropic: @anthropic-ai/claude-agent-sdk, @anthropic-ai/sdk
• MCP SDK: @modelcontextprotocol/sdk (functions/, agents/store-mcp)
• Cloudflare: D1 + R2 + cloudflared (the platform's persistence + ingress)
• Testing: vitest (the platform's e2e tiers)

Authentication

Claude API auth, used across projects:

1. OAuth Token (Claude Pro/Max): npm install -g @anthropic-ai/claude-cli && claude setup-token
2. API Key (developers): from console.anthropic.com, into the project's .env as ANTHROPIC_API_KEY

The platform also uses a scoped Cloudflare R2 API token (in agents/.env); see agents/docs/r2-setup.md.

Testing Philosophy

Real tests only — no mocks, no stubs. If it doesn't exercise actual behavior, we don't do it.

• agents/: real child-process servers over real A2A — fast (npm run test:e2e, CI), live (npm run test:live, real engine/browsers/R2), soak (npm run test:soak, 10× loop)
• functions/: E2E with real Anthropic + da.live APIs
• content-authoring-eval/: manual via web UI + curl
• agent-claude-sdk/ / make-dot-com/: ad-hoc / in-platform

Documentation Standards

Each subproject carries a README.md (user-facing: quick start, features, links) and a CLAUDE.md (AI context: architecture, gotchas, workflows). The agents/ platform additionally gives each workspace its own CLAUDE.md and ships an as-built build report under ai-docs/.

Related Resources

• A2A Protocol · Claude Agent SDK · Anthropic API · MCP Spec · da.live Admin API · Cloudflare Developers

License

Apache License 2.0

---

Last Updated: 2026-08-29 Primary Tools: Claude Code, A2A SDK, Agent SDK, Azure Functions, Next.js, Cloudflare (D1/R2/Tunnel), MCP, opencode (Kimi K3/K2.7 - optional since 2026-10, no subscription) Active workstream: agents/ v2.x platform (v2.8.1) → the Wilderness Journal demo site (adapt-to-2026-demo) for adaptTo() Sept 2026. content-authoring-eval/ + agent-claude-sdk/ are deprecated.