aep

Adobe's MCP lets your agent read Experience Platform. This one lets it work.

51 tools across 12 categories. Full read AND write. Self-hosted, Apache-2.0, no invitation required. Batch ingestion, schema composition, audience activation, data lifecycle, privacy jobs, query service.

Ingest a batch → compose a schema → activate an audience → honour an erasure. Every mutation gated by a fail-closed write guard that asks Adobe what kind of sandbox it's in.

⚡ Do it in three lines

claude mcp add aep \
  -e AEP_CLIENT_ID=... -e AEP_CLIENT_SECRET=... \
  -e AEP_ORG_ID=...@AdobeOrg -e AEP_SANDBOX_NAME=your-dev-sandbox \
  -- npx -y @focusgts/aep-mcp-server

Then just ask your agent:

"Create a schema with the Demographic Details field group, then a dataset on it." "Ingest this NDJSON file and tell me when the batch lands." "Build an audience of customers who bought twice this quarter and activate it." "Delete every record for this email address — dry run first."

Writes are off until you ask for them, and safe mode only unlocks sandboxes Adobe classifies as development.

The loop that makes it different

flowchart LR
  A["📐 Composeschema from field groups"] --> B["🗂️ Createdataset"]
  B --> C["📥 Ingestbatch · upload · complete"]
  C --> D["🎯 Activatesegment → destination"]
  D --> E["🧹 Governerasure · expiration · quota"]
  E -. "re-audit the tenant" .-> A

Adobe's first-party gateway can tell you what's in your Experience Platform tenant. It cannot create a dataset, land a batch, activate an audience, or submit an erasure. This does — and does it behind a guard that fails closed.

🧠 How it works

flowchart LR
  A["AI agent(Claude · Cursor · Copilot)"] -- MCP / stdio --> B["aep-mcp-server51 tools"]
  B --> W{{"write guardfail-closed"}}
  W --> C["Schema Registry+ Catalog"]
  W --> D["Batch Ingestion"]
  W --> E["Profiles · SegmentsDestinations"]
  W --> G["Data Lifecycle+ Privacy"]
  C --> F["Your AEP sandbox"]
  D --> F
  E --> F
  G --> F

The agent calls tools; the server talks to live Adobe APIs over OAuth Server-to-Server. The write guard sits in the HTTP client, not in each tool, so all 51 inherit it and none can forget it. Blocked calls never reach Adobe.

🛡️ Safe by default

Three postures. Reads are never restricted in any mode.

How safe decides — and why it's not the sandbox name. A production sandbox can be called anything, and a sandbox called prod might not be production. Only Adobe's type field from the Sandbox Management API decides.

It fails closed. If the type can't be determined — the credential can't read sandbox metadata, the API errors, startup hasn't finished — writes are blocked. A credential must not earn write access by being less capable. An unrecognised AEP_MODE falls back to safe, so a typo can never grant production writes.

A sandbox literally named prod is refused unconditionally, before mode resolution — so AEP_MODE=production does not lift it. Override with AEP_I_UNDERSTAND_THIS_WRITES_TO_PROD=true only if that really is your sandbox's name. The inference is deliberately asymmetric: trusting a name to allow a write is unsafe, trusting one to deny a write is safe, because the worst case is a refusal you can override on purpose.

Mutations are off entirely unless AEP_ALLOW_MUTATIONS=true. That is separate from AEP_MODE on purpose: choosing a write mode should not also mean "yes, you may change my data".

Startup always states the active posture:

SAFE MODE — sandbox is a development sandbox, so writes are ENABLED.
SAFE MODE — sandbox is PRODUCTION, so writes are BLOCKED. Reads work normally.
SAFE MODE — sandbox type could not be confirmed, so writes are BLOCKED (fail-closed).
READ-ONLY MODE — no write, update, or delete will be performed in any sandbox.
PRODUCTION MODE — writes permitted against ANY sandbox, including production.

Per-tool confirmation gates

Writes are not uniformly gated — uniform gating makes an agent useless. Gates sit where an action is irreversible and wide-reaching, and every one is checked before any network call:

Confirmations name their target. A phrase carrying the dataset id — and for record delete, a hash of the identities too — cannot be copied from one call to another. A generic "I understand this is irreversible" approves any deletion once you've typed it once.

Identity values never leave the process. aep_create_record_delete returns a count, the namespace names, and a digest — never the email addresses or device IDs you passed it. A record-delete request is by nature a list of real people; a tool that echoes them copies them into every transcript and log sink it touches.

Batch creation and file upload are ungated on purpose: those writes are additive and recoverable. An unwanted batch can be left uncompleted, and data that did land can be removed with the Data Hygiene tools.

Tool annotations

Every tool ships MCP annotations — readOnlyHint, destructiveHint, idempotentHint, openWorldHint — derived from the same metadata that builds its description, so the two cannot drift.

These are hints for the client, not enforcement — the guards above enforce. Their value is that a client like Claude Desktop uses destructiveHint to decide when to interrupt and ask a human. Without them aep_delete_profile looks identical to aep_list_schemas. A test asserts the destructive list exactly, so a ninth is a deliberate act rather than an oversight.

🛠️ The 51 tools

All prefixed aep_, verb_noun naming. 🔒 changes state · 🔥 destructive.

Data modelling & ingestion

Schemas (4)

• list_schemas
• get_schema
• create_schema 🔒
• update_schema 🔒

Datasets (4)

• list_datasets
• get_dataset
• create_dataset 🔒
• delete_dataset 🔥

Ingestion (7)

• create_batch 🔒
• upload_batch_file 🔒
• complete_batch 🔥
• get_batch_status
• list_batches
• abort_batch 🔥
• revert_batch 🔥

Sources (2)

• list_sources
• list_dataflows

Query Service (3)

• run_query 🔒
• get_query_status
• list_queries

Profiles, audiences & activation

Identities (2)

• list_identity_namespaces
• get_identity_graph

Profiles (4)

• get_profile
• get_profile_by_identity
• preview_profile
• delete_profile 🔥

Segments (5)

• list_segments
• get_segment
• create_segment 🔒
• estimate_segment_size
• delete_segment 🔥

Destinations (3)

• list_destinations
• create_destination_connection 🔒
• activate_segment 🔒

Governance — privacy, lifecycle & event routing

Privacy Service (6)

• create_privacy_job 🔒
• get_privacy_job
• list_privacy_jobs
• cancel_privacy_job 🔒
• get_privacy_job_results
• list_privacy_namespaces

Data Hygiene (9)

• create_record_delete 🔥
• get_work_order_status
• list_work_orders
• get_data_lifecycle_quota
• create_dataset_expiration 🔥
• get_dataset_expiration
• list_dataset_expirations
• update_dataset_expiration 🔥
• cancel_dataset_expiration 🔥

Workflows these unlock

Ingest end to end — create_schema → create_dataset → create_batch → upload_batch_file → complete_batch → get_batch_status

Build and activate an audience — create_segment → estimate_segment_size → list_destinations → create_destination_connection → activate_segment

Honour an erasure request — get_profile_by_identity → create_record_delete → get_work_order_status

Retire data on a schedule — create_dataset_expiration → list_dataset_expirations → update_dataset_expiration → cancel_dataset_expiration

What's actually been run against a live tenant is recorded per tool in docs/VALIDATION-MATRIX.md — including the surfaces that are documented-and-mocked but deliberately never executed, and why.

Adobe Journey Optimizer (2)

AJO is a separate Adobe product, licensed separately — hence the ajo_ prefix, so an entitlement failure reads as one.

Campaigns

• ajo_list_campaigns
• ajo_get_campaign

Campaigns is the only AJO surface reachable on our tenant. Journeys, messages, channel surfaces, content templates, fragments, offers and decisions all return an HTML 404 — the gateway has no such route — so they are deliberately not implemented.

Writes are absent on purpose. The routes exist, but shipping an unvalidated write path into a product that sends messages to real people is not a trade worth making.

📊 AEC-Bench — does your agent actually work?

Every MCP server in this space is described by its tool count. That measures surface area, not competence: fifty tools that 404 score higher than ten that work.

bench/ is an agentic benchmark that measures the other thing — given a real task and a live tenant, does the agent finish it, and can you prove it?

npm run bench          # tier 1, read-only, safe on any tenant
npm run bench:write    # tier 2, creates and removes what it creates

Current: tier 1 5/5, tier 2 2/2, zero residue. Tier 3 (irreversible) is defined and deliberately empty — its tasks are non-cancellable and can take 30 days, and a benchmark is not a good reason to run one.

We expect to score badly on tasks we haven't built for. That's the intended use.

🥊 vs Adobe's first-party Experience Platform tools

Adobe ships first-party tools through CX Coworker Gateway. It's a genuinely good product, and if all you need is to ask questions about your tenant, use it.

Audiences and destinations do appear, but in a separate Real-Time CDP tool set on the same gateway — and Adobe is explicit that creating, activating, updating, or deleting audiences, destinations, and dataflows isn't supported there either. The read-only boundary holds across the whole gateway.

They're complementary, not competing: pair Adobe's gateway for governed reads with this server for the write path.

Adobe's figures were read from their Experience Platform tools page (last updated 17 July 2026): search_datasets, search_class_relations, search_data_access, search_data_lake, search_dule, search_query_service, search_audit, search_allowed_ip_ranges. It's a Beta surface and will change — check their docs for the current figure. The read/write split is the durable difference, not the count.

🔌 Add it to your tool

Claude Code — one command

claude mcp add aep \
  -e AEP_CLIENT_ID=... -e AEP_CLIENT_SECRET=... \
  -e AEP_ORG_ID=...@AdobeOrg -e AEP_SANDBOX_NAME=your-dev-sandbox \
  -- npx -y @focusgts/aep-mcp-server

Claude Desktop — claude_desktop_config.json

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "aep": {
      "command": "npx",
      "args": ["-y", "@focusgts/aep-mcp-server"],
      "env": {
        "AEP_CLIENT_ID": "...",
        "AEP_CLIENT_SECRET": "...",
        "AEP_ORG_ID": "...@AdobeOrg",
        "AEP_SANDBOX_NAME": "your-dev-sandbox"
      }
    }
  }
}

Cursor — .cursor/mcp.json

{
  "mcpServers": {
    "aep": {
      "command": "npx",
      "args": ["-y", "@focusgts/aep-mcp-server"],
      "env": {
        "AEP_CLIENT_ID": "...",
        "AEP_CLIENT_SECRET": "...",
        "AEP_ORG_ID": "...@AdobeOrg",
        "AEP_SANDBOX_NAME": "your-dev-sandbox"
      }
    }
  }
}

VS Code (GitHub Copilot) — .vscode/mcp.json

{
  "servers": {
    "aep": {
      "command": "npx",
      "args": ["-y", "@focusgts/aep-mcp-server"],
      "env": {
        "AEP_CLIENT_ID": "...",
        "AEP_CLIENT_SECRET": "...",
        "AEP_ORG_ID": "...@AdobeOrg",
        "AEP_SANDBOX_NAME": "your-dev-sandbox"
      }
    }
  }
}

🔐 Credentials

Get them at developer.adobe.com/console: create a project → add Experience Platform API → OAuth Server-to-Server.

AEP_SANDBOX_NAME has no default, deliberately. It used to fall back to prod, which meant a config file missing one line silently pointed every request — reads included — at production, with no warning. There is no safe default: a wrong guess is indistinguishable from a correct one until something is read or written in the wrong environment. Setting it explicitly to prod is allowed; that's a visible, deliberate choice, and mutations there are still refused by the write guard.

Sandbox scoping. Every tool sends x-sandbox-name, and Query Service derives its database as <AEP_SANDBOX_NAME>:all.

🧾 Entitlements

Not every Adobe org licenses every AEP product. A tool returning AEP_403 usually means a missing entitlement rather than a bad credential.

🏗️ Architecture

TypeScript strict end-to-end, @modelcontextprotocol/sdk + zod, stdio transport, stateless per request.

flowchart LR
  C["MCP clientClaude · CursorCopilot · ChatGPT"]
  T["aep-mcp-server51 tools12 categories"]
  G{{"write guardfail-closed"}}
  A1["Schema Registry· Catalog"]
  A2["Batch Ingestion"]
  A3["UPS · Segmentation· Destinations"]
  A4["Data Lifecycle· Privacy"]
  IMS[/"Adobe IMSOAuth S2S"/]

  C -- "stdio · JSON-RPC 2.0" --> T
  T -- "every call, no exceptions" --> G