ExD Accelerator — MCP Server

AI-native Experience Decisioning lifecycle automation for Adobe Journey Optimizer. End-to-end ExD setup from a single chat conversation: CSV → schema fields → offers → collections → eligibility rules → ranking → selection strategy → placements.

21 MCP tools wrapping AEP Schema Registry and Decisioning APIs. Every write operation previews what it will do and requires explicit confirmed: true before executing.

Live endpoint: https://exd-mcp-server-without-auth.vercel.app/api/mcp Health: https://exd-mcp-server-without-auth.vercel.app/api/health

---

What's new in 2.0

Was in 1.x	Now in 2.0
Manual ACCESS_TOKEN paste, expires every 24h	Auto-mints via client_credentials, cached per client_id
Express + SSE local server, ngrok tunnel for sharing	Streamable HTTP serverless route — deploy to Vercel in one click
Single hardcoded sandbox in .env	Per-request config via HTTP headers — every coworker uses their own sandbox
README claimed 13 tools, code had 21	21 tools, documented
Bulk offers defaulted to 2024-06-10 (past date)	Defaults to today + 1 year
lookup_decisioning_schema showed 0 fieldgroups on healthy schemas	Reads from meta:extends too — accurate counts
.gitignore had ".env" (quoted) — would not actually ignore	Plain .env, plus .vercel/, log files, etc.

---

Two ways to run this

A. Local stdio (Claude Desktop)

npm install
cp .env.example .env       # fill in CLIENT_ID, CLIENT_SECRET, sandbox, schema, catalog
npm start                  # runs src/stdio.js

Then point Claude Desktop at it (%APPDATA%\Claude\claude_desktop_config.json on Windows, ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "exd-accelerator": {
      "command": "node",
      "args": ["/absolute/path/to/exd-mcp-server/src/stdio.js"],
      "env": {
        "CLIENT_ID":                 "…",
        "CLIENT_SECRET":             "…",
        "ORG_ID":                    "…@AdobeOrg",
        "SANDBOX_NAME":              "…",
        "TENANT_ID":                 "…",
        "DECISIONING_SCHEMA_URI":    "https://ns.adobe.com/…/schemas/…",
        "DECISIONING_SCHEMA_ALT_ID": "_….schemas.…",
        "ITEM_CATALOG_ID":           "xcore:decision-catalog:…"
      }
    }
  }
}

Restart Claude Desktop. You'll see the 🔧 tool icon — ExD Accelerator is live.

B. Vercel deployment (for sharing with Adobe coworkers)

1. Push this repo to GitHub.
2. Import it in vercel.com/new. Framework preset: Other.
3. Deploy — Vercel auto-detects api/mcp.js as the serverless route.
4. Do not add Adobe credentials to Vercel environment variables. Each coworker supplies their own credentials via HTTP headers when they connect.

Adobe coworker setup

Once deployed at https://your-app.vercel.app, each coworker adds this in their MCP client config (Claude Desktop, Adobe AO Chat, Claude.ai, etc.):

Setting	Value
Server URL	https://your-app.vercel.app/api/mcp
Transport	Streamable HTTP
Auth	None at the transport level — credentials go in headers

Custom headers (one-time, in the MCP client config — NOT in chat messages):

x-adobe-client-id:      <Adobe Dev Console: Client ID>
x-adobe-client-secret:  <Adobe Dev Console: Client Secret>
x-adobe-org-id:         <IMS Org ID>@AdobeOrg
x-adobe-sandbox:        <sandbox name>
x-adobe-tenant-id:      <tenant id, e.g. acssandboxgdcthree>
x-adobe-schema-uri:     https://ns.adobe.com/<tenant>/schemas/<id>
x-adobe-schema-alt-id:  _<tenant>.schemas.<id>
x-adobe-catalog-id:     xcore:decision-catalog:<id>

The server uses client_credentials to mint a token automatically, caches it per client_id, and refreshes on expiry. The marketer never sees the token.

---

Is Vercel a good fit for a production MCP endpoint?

Yes, for this workload. Each MCP tool call is a single short HTTP roundtrip to Adobe Platform APIs — no long-running state, no streaming required, no WebSocket. Vercel's serverless model maps cleanly:

Concern	Verdict
Stateless requests	✅ Each MCP call is independent. No session state.
Cold-start latency	⚠️ ~300–500ms on first call after idle. Subsequent calls reuse the warm container.
60s function timeout (Pro tier)	⚠️ bulk_create_offers with >50 rows may exceed this. Chunk large batches.
10s timeout (Hobby tier)	⚠️ get_setup_summary is fine; large bulk is not. Upgrade to Pro for production use.
Auto-scaling	✅ Each coworker's request gets its own invocation.
HTTPS, custom domain, env-per-deploy	✅ Built in.
SSE / long-polling	❌ Not used here — we run Streamable HTTP with enableJsonResponse: true, which is single-request/response and fits serverless perfectly.

When Vercel isn't right: if you need server-initiated notifications, very large bulk operations (hundreds of writes), or stateful sessions across many calls, deploy to a long-lived host (Railway, Render, Fly, ECS) and run the same codebase. The transport layer is the only difference.

---

All 21 tools

Read-only (no confirmation needed)

#	Tool	What it does
1	parse_csv_and_suggest	Parses CSV, infers XDM types per column, suggests eligibility rules and ranking formulas. Always call first. No API calls.
9	get_offer_item	Fetches a single offer item by DPS ID
10	list_offer_items	Lists all offers in catalog with pagination
16	get_setup_summary	Full inventory: offers, collections, rules, formulas, strategies, placements
17	lookup_decisioning_schema	Full resolved schema with OOB + tenant fields; accepts include_deprecated: true
18	list_schema_fieldgroups	Lists all tenant fieldgroups for the offer item class
19	get_fieldgroup	Full field definitions inside a specific fieldgroup
20	get_schema_audit_log	Chronological change history for the decisioning schema
21	list_schema_descriptors	Identity, deprecation, display name, relationship descriptors

Write (require confirmed: true)

#	Tool	What it does
2	create_offer_metadata_fieldgroup	Creates XDM fieldgroup from CSV columns, attaches to decisioning schema. Checks for duplicates first.
3	bulk_create_offers	Creates offer items from CSV rows. Supports dry_run: true for payload preview
4	create_collection	Creates offer collection with filter constraint
5	create_eligibility_rule	Creates PQL eligibility rule
6	create_ranking_formula	Creates ranking formula (static, custom field, recency-hybrid, custom PQL)
7	create_selection_strategy	Wires collection + rule + formula into a selection strategy
8	create_placement	Creates channel placement via /exd-placements endpoint
11	update_offer_item	JSON Patch update on any offer field
12	add_schema_field	Adds a single field to an existing tenant fieldgroup
13	deprecate_schema_field	Sets meta:status: deprecated on a custom tenant field
14	deprecate_oob_field	Creates xdm:descriptorDeprecated for OOB Adobe-managed fields
15	detach_fieldgroup	Removes fieldgroup from schema allOf and meta:extends

Confirmation pattern

Every write tool shows a preview and blocks with:

⚠️  CONFIRMATION REQUIRED — no changes made yet
[preview of what will happen]
✅ To proceed, call this tool again with confirmed: true

Call the same tool again with confirmed: true to execute.

---

Recommended workflow from a fresh CSV

1.  parse_csv_and_suggest            → analyse CSV, no writes
2.  list_schema_fieldgroups          → check if fieldgroup already exists
3.  create_offer_metadata_fieldgroup → push schema fields (confirmed: true)
4.  lookup_decisioning_schema        → verify fields attached
5.  bulk_create_offers (dry_run)     → preview offer payloads
6.  bulk_create_offers (confirmed)   → create offers
7.  list_offer_items                 → verify
8.  create_collection                → group offers (confirmed: true)
9.  create_eligibility_rule          → targeting (confirmed: true)
10. create_ranking_formula           → ranking logic (confirmed: true)
11. create_selection_strategy        → wire it all together (confirmed: true)
12. create_placement                 → define channel (confirmed: true)
13. get_setup_summary                → verify full setup

---

Sample CSV for testing

name,description,category,brand,discount_percent,price,region,priority,start_date,end_date
Summer Glow Kit,Complete summer skincare set,Skincare,GlowCo,20,49.99,US,1,2024-06-01,2024-08-31
SPF Starter Bundle,SPF 30 and 50 combo,Skincare,GlowCo,15,29.99,US,2,2024-06-01,2024-09-30
Loyalty 20% Off,Exclusive 20% for gold members,Discount,GlowCo,20,0,Global,1,2024-01-01,2024-12-31

Column mapping:

• name → itemName (OOB), description → itemDescription, priority → itemPriority, start_date/end_date → itemCalendarConstraints
• everything else → _<tenant>.<column> (custom fieldgroup)

---

File layout

exd-mcp-server-without-auth/
├── src/
│   ├── server.js         ← buildMcpServer(config) + 21 tool definitions
│   ├── stdio.js          ← stdio entry (npm start) — for Claude Desktop
│   └── http-local.js     ← local HTTP server for testing the Vercel route
├── api/
│   └── mcp.js            ← Vercel serverless route (Streamable HTTP)
├── scripts/
│   └── smoke.js          ← smoke test for stdio + HTTP transports
├── vercel.json
├── package.json
├── .env.example
└── .gitignore

---

Smoke testing

npm install
cp .env.example .env       # fill in
npm run smoke              # runs stdio + HTTP transport tests, calls real Adobe APIs

Expected output ends with All smoke checks passed.

---

Troubleshooting

Symptom	Likely cause	Fix
Missing credentials	CLIENT_ID/CLIENT_SECRET not set	Add to .env (local) or to your MCP client's header config (deployed)
IMS token mint failed (401)	Credentials invalid or revoked	Regenerate the OAuth Server-to-Server credential in Adobe Developer Console
401 Oauth token is not valid from Adobe	Credential lacks AEP access	The OAuth credential's product profile needs Adobe Experience Platform access for the target sandbox
403 Forbidden	Wrong org/sandbox	Check ORG_ID and SANDBOX_NAME
List offers returns 0	Wrong ITEM_CATALOG_ID for the sandbox	Each sandbox has its own catalog ID
Tool call exceeds 10s on Vercel Hobby	Bulk operation too large	Upgrade to Pro (60s) or chunk the CSV
lookup_decisioning_schema shows 0 fieldgroups	Resolved by 2.0 — file an issue if you still see this	—

---

What this MCP does NOT do (future scope)

• Decisioning policy / campaign creation — creates components but not the final AJO policy that ties strategy + placement.
• Delete operations — AEP recommends archive over delete.
• Audience creation — eligibility rules reference profile attributes but don't create AEP segments.
• Cross-channel coherence scoring — would require AEP Query Service integration.