Mcp Server Sfmc

mcp-server-sfmc

MCP server providing Salesforce Marketing Cloud language intelligence — AMPscript, SSJS, GTL, and Marketing Cloud Next (MCN) Handlebars — as Model Context Protocol tools, resources, and prompts for AI-assisted development and code review. Ships two searchable doc indexes: MCE help (Salesforce Help for Marketing Cloud Engagement administration and setup) and MCN developer docs (Marketing Cloud Next REST API and developer reference). Includes full Marketing Cloud Next (MCN) platform awareness: auto-detects whether a project targets MCE or MCN, restricts completions and validation to MCN-supported functions, flags SSJS as unsupported in MCN, surfaces MCN behavioral differences in hover and diagnostics, validates MCN Handlebars templates, and provides a migration toolkit to analyze and rewrite code for MCN — including AMPscript ↔ Handlebars and SSJS → Handlebars conversion.

Built on sfmc-language-lsp, the same engine that powers the SFMC Language Service VS Code extension.

Shared-diagnostic migration

sfmc-language-lsp 4.0.0 changes shared validator codes to canonical sfmc/... IDs and preserves quick-fix variants and original payloads in a namespaced envelope. See the LSP migration and rule documentation. Documentation URLs are owned by the LSP version, not this MCP package.

Consumer regression coverage checks the canonical SSJS diagnostics, distinct polyfill/replacement payloads, and overlap suppression. This is not a new MCP tool or a promise that every tool response exposes raw LSP diagnostics.

Version 2.10.0 adopts ampscript-data: ^4.2.0 and sfmc-language-lsp: ^4.2.0 from npm. This Winter ’27 (API 68) catalog update makes the current AMPscript and shared language-service metadata available to MCP clients. MCP validation and review output continues to format locations, severities, and messages rather than expose raw LSP codes or envelopes.

VS Code MCP Server Gallery (@mcp)

This package is registered with the official MCP Registry as io.github.JoernBerkefeld/mcp-server-sfmc so it can appear in Visual Studio Code when you use the @mcp filter in the Extensions view (see the publish quickstart). Enable chat.mcp.gallery.enabled if the gallery does not show.

The registry only stores metadata; the server still runs locally via stdio (for example npx -y mcp-server-sfmc@latest). This is separate from @contribute:mcp, which lists VS Code extensions that contribute MCP definitions — use the SFMC Language Service for that path.

After publishing metadata (see Publish an MCP Server or the release workflow), you can confirm the entry with:

curl "https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.JoernBerkefeld/mcp-server-sfmc"

VS Code without manual MCP config

If you use the SFMC Language Service extension (1.101+), it registers this MCP server for discovery in VS Code — you normally do not need to edit .vscode/mcp.json or run npm install for that path; VS Code still launches the published package via npx when the server starts.

For other editors, or if you prefer explicit configuration, use the npx or install options below.

Using this package without the VS Code extension

You do not have to install the VS Code extension. Pick one way to run the server:

None of these replace the VS Code extension for editing (syntax, LSP, snippets); they only expose the MCP server to tools that speak the Model Context Protocol.

CI templates and sfmc-review-diff

Ready-to-copy workflows (GitHub Actions, GitLab CI, Jenkins, Azure Pipelines, Bitbucket Pipelines) live under ci-templates/. They run eslint-plugin-sfmc on changed files and sfmc-review-diff on the PR/MR unified diff. See ci-templates/README.md for what each file does and how the two checks differ.

What it gives your AI assistant

Mixed HTML validation boundary

Starting with version 2.10.1, validate_sfmc_html validates embedded AMPscript/GTL and applies SSJS checks only to actual server JavaScript bodies (runat="server", excluding AMPscript script bodies). Browser scripts and non-JavaScript script data are not sent to the SSJS validator. Server diagnostics retain locations in the original HTML, and browser declarations, Core loads, or polyfills cannot suppress server warnings. HTML without server JavaScript does not receive an SSJS incompatibility error solely because target is 'next'.

This tool does not validate browser JavaScript or provide DOM/type diagnostics. The standalone validate_ssjs tool continues to validate raw SSJS with its existing target-platform behavior.

This boundary uses the published validateMixedDocumentSsjs API from sfmc-language-lsp 4.3.1 or newer. Version 2.10.1 raises the dependency floor to ^4.3.1; release verification uses a clean registry-only installation.

Connecting AI clients

VS Code (1.99+) + GitHub Copilot — manual mcp.json

If you are not using the SFMC Language Service extension's built-in MCP registration, add a .vscode/mcp.json file to your project (or copy it from this repo):

{  "servers": {    "sfmc": {      "type": "stdio",      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    }  }}

Open the file in VS Code — a Start button appears at the top. Click it to launch the server. Open GitHub Copilot Chat in Agent mode and the SFMC tools appear automatically.

Cursor

Add to your Cursor settings (~/.cursor/mcp.json or project-level .cursor/mcp.json):

{  "mcpServers": {    "sfmc": {      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    }  }}

Restart Cursor. The tools are available in Agent mode.

Claude Desktop

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

{  "mcpServers": {    "sfmc": {      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    }  }}

Restart Claude Desktop.

Windsurf

Add to your Windsurf MCP settings (~/.codeium/windsurf/mcp_config.json):

{  "mcpServers": {    "sfmc": {      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    }  }}

Local install (faster startup than npx)

npm install -g mcp-server-sfmc

Then replace "command": "npx", "args": ["-y", "mcp-server-sfmc@latest"] with:

"command": "mcp-server-sfmc","args": []

Tools

Resources

Prompts

Access via /mcp.sfmc.writeAmpscript etc. in VS Code, or via the prompts API:

Migrating code to Marketing Cloud Next

Marketing Cloud Next (MCN) supports 41 of the 155 AMPscript functions and does not support SSJS. Three supported functions have behavioral differences (see below). The migration toolkit guides you from analysis through to a fully rewritten result.

Recommended workflow

1. detect_sfmc_platform  →  confirm the project targets MCN2. check_mcn_compatibility  →  understand scope: which files need work, how hard3. rewrite_for_mcn prompt  →  calls the rewrite_for_mcn tool internally, then applies AI                               reasoning to any MANUAL_REWRITE_REQUIRED sections4. validate_ampscript with target:'next'  →  verify the rewritten code is clean

What check_mcn_compatibility tells you

Pass one or more { filename, content } pairs and get back:

• Per-function classification for every AMPscript call site:
  • ✅ Supported — works as-is
  • ⚠️ Needs review — supported but has behavioral differences in MCN
  • ❌ Not supported — no MCN equivalent
• SSJS block analysis — classified as "needs conversion" (has AMPscript equivalents) or "not migratable" (uses JS-native constructs such as try/catch, forEach, regex)
• Migration difficulty per file — Ready / Minor changes needed / Significant rewrite / Not migratable
• Executive summary across all files

What rewrite_for_mcn does

Use the rewrite_for_mcn prompt — it is the primary interface and gives you the full result in one step. It first calls the rewrite_for_mcn tool internally for all deterministic rewrites, then applies AI reasoning to every section the tool could not handle mechanically. You get both the accuracy of rule-based conversion and the intelligence of an AI pass, without having to invoke them separately.

Deterministic rewrites (handled by the tool internally):

• Removes StringToDate() wrappers from FormatDate() calls (FormatDate(StringToDate(@x), fmt) → FormatDate(@x, fmt))
• Converts .NET → Java SimpleDateFormat format strings (e.g. tt → a)
• Annotates Lookup() calls with an odd number of search arguments
• Marks MCE-only AMPscript functions with %%-- NOT SUPPORTED IN MCN %%
• Converts convertible SSJS blocks to AMPscript (Platform.Function.* → AMPscript equivalents, Platform.Variable.GetValue/SetValue → @variable references)
• Flags JS-native SSJS constructs (try/catch, array methods, regex, complex logic) as MANUAL_REWRITE_REQUIRED

AI reasoning layer (added by the prompt): for each MANUAL_REWRITE_REQUIRED section, the prompt instructs the AI to attempt a full conversion using its knowledge of AMPscript and SSJS, then produces a single final rewritten code block with a prose changelog covering every change.

Pipelines and programmatic use: call the rewrite_for_mcn tool directly when you need the structured JSON output ({ rewrittenCode, changes[], nonMigratableItems[], difficulty, summary }) without the AI layer — for example in a CI step that feeds results to another tool.

The same hybrid pattern applies to convertSsjsToAmpscript and convertAmpscriptToSsjs: always use the prompt for interactive use; use the tool when you need structured output in a pipeline.

MCN behavioral differences to watch for

CloudPages functions — not migratable

Functions that depend on CloudPages context (CloudPagesURL, RequestParameter, QueryParameter, Redirect, MicrositeURL) have no equivalent in Marketing Cloud Next. check_mcn_compatibility and rewrite_for_mcn flag these as Not migratable immediately.

MCN Handlebars

Marketing Cloud Next embeds a locked-down Handlebars templating layer (Handlebars.Net) alongside AMPscript. The server exposes the full helper catalog and conversion tooling, all data-driven from handlebars-data and the handlebarsEquivalent / handlebarsExact fields in ampscript-data — nothing is hand-maintained.

Authoring & lookup

1. list_handlebars_helpers       →  browse the catalog (filter by category / origin)2. lookup_handlebars_helper      →  signature, parameters, origin, MCN API version3. get_handlebars_completions    →  prefix-filtered completions4. writeHandlebars prompt        →  generate Handlebars-in-HTML, validated against the catalog5. validate_handlebars           →  helper names, arity, block balance, unsupported constructs

validate_handlebars rejects constructs the MCN engine does not implement — partials ({{> name}}), decorators ({{#* ...}}), and built-in helpers absent from the locked-down runtime.

Conversion categories

AMPscript → Handlebars conversion classifies every function from ampscript-data:

The Approximate note names the closest helper so consumers can tell a function whose call shape merely differs (e.g. ContentBlockByKey → getContentBlock) apart from one with no MCN counterpart at all. convertSsjsToHandlebars is transitive (SSJS → AMPscript → Handlebars) and conservatively flags imperative SSJS for manual rewrite.

Writing effective prompts

Automatic tool use

Clients that honour the MCP instructions field (Cursor, Claude Desktop, GitHub Copilot Agent mode) will call search_help or search_mce_help automatically whenever you ask an MCE administration or setup question — no special phrasing needed. If your client does not process server instructions, or if you want explicit control, the templates below help.

Quick reference: which tool or prompt to use

Copy-paste prompt templates

Classic Engagement admin (most common)

Search the Marketing Cloud Engagement help (product_focus: engagement) and tell me:<your question here>
Cite which product version your steps apply to and note if the bundled docs are incomplete.

Marketing Cloud Next — developer API

Search the Marketing Cloud Next developer docs and tell me:<your question here>
Confirm the steps apply to Marketing Cloud Next, not classic Engagement.

MCN migration analysis

Check these files for Marketing Cloud Next compatibility and give me an executive summary:<paste file contents>

Rewrite code for MCN

Use the rewrite_for_mcn prompt — it first runs the deterministic rewrite tool, then applies AI reasoning to any sections that need it. You get the full result in one step.

Rewrite the following AMPscript/HTML for Marketing Cloud Next:<paste code>
Apply all deterministic fixes, convert convertible SSJS to AMPscript, and explain every change.

Unknown product / migration question

Search both Marketing Cloud Engagement and Next help (product_focus: any) and tell me:<your question here>
Separate the steps for classic Engagement vs Marketing Cloud Next clearly.

Explicit use of the answerMceHowTo prompt

In clients that support MCP prompts (e.g. VS Code with the /mcp.sfmc.answerMceHowTo command):

/mcp.sfmc.answerMceHowTo  question: "How do I create a child business unit and assign it a sender profile?"  assumeProduct: engagement

Combining MCE admin with code

1. Use search_mce_help (product_focus: engagement) to find the correct Journey Builder entry event   configuration steps, then2. Write the AMPscript snippet that reads the data extension row inside the journey email.

What to expect from the AI

• The AI cites the product scope (Engagement or Next) in every answer.
• If the bundled excerpts do not cover the question fully, the AI says so and suggests verifying in the live Salesforce Help.
• Function signatures (AMPscript / SSJS) are always sourced from the language catalog, not from training data.

Refresh bundled help indexes

Marketing Cloud Engagement help

The published npm package includes bundled/mce-help/chunks.json, produced by npm run bundle-mce-help. By default the bundler uses the path in MCE_HELP_DOCS (see scripts/bundle-mce-help.mjs); set it to the root of your mirrored MCE Help Markdown tree when not using the maintainer default.

cd mcp-server-sfmcnpm run bundle-mce-helpnpm run buildnpm test

Example: MCE_HELP_DOCS=/absolute/path/to/mce-help-mirror npm run bundle-mce-help

Marketing Cloud Next developer docs

The published npm package includes bundled/mcn-help/chunks.json, produced by npm run bundle-mcn-help. Set MCN_HELP_DOCS to the root of your mirrored MCN developer docs tree when not using the maintainer default (see scripts/bundle-mcn-help.mjs).

cd mcp-server-sfmcnpm run bundle-mcn-helpnpm run buildnpm test

To rebuild both indexes in one command: npm run bundle-all

AI code review in pull requests

Cursor Bugbot

Cursor Bugbot reviews pull requests (GitHub) and merge requests (GitLab) automatically, leaves inline comments with fix suggestions, and publishes a Cursor Bugbot check. It does not run sfmc-review-diff itself — you make it SFMC-aware in two ways:

1. Repository rules (BUGBOT.md) — all plans. Copy ci-templates/BUGBOT.md to .cursor/BUGBOT.md in your repo root. It tells Bugbot how to review AMPscript, SSJS, and SFMC HTML — unknown functions, arity, delimiter balance, ES3-only SSJS, and optional Marketing Cloud Next migration rules. Bugbot reads the root file plus any nested .cursor/BUGBOT.md files near changed files, so you can scope stricter rules to subfolders.

2. MCP tools (mcp-server-sfmc) — Team / Enterprise plans only. Bugbot can call MCP tools during a review so its findings come from the same language catalog as the editor:

1. In the Bugbot dashboard, enable Bugbot on the repository.
2. Open the MCP configuration for Bugbot and add this server:

{  "mcpServers": {    "sfmc": {      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    }  }}

3. Add the tools to Bugbot in the dashboard. The relevant review tools are review_change, validate_ampscript, validate_ssjs, validate_sfmc_html, and suggest_fix. The BUGBOT.md rules above instruct Bugbot when to call them.

Branch protection: require the Cursor Bugbot check to make sure a review runs before merge. Findings default to a neutral conclusion; enable fail-on-unresolved-issues in the dashboard (where available) if you want findings to produce a failing check.

Run it locally first. In Cursor 3.7+, run /review-bugbot on your branch before you push; Bugbot reuses that review on the PR/MR with the same diff, avoiding a duplicate run.

GitHub Copilot (cloud agent)

The .github/agents/sfmc-reviewer.agent.md custom agent in this repository configures a GitHub Copilot cloud agent that uses mcp-server-sfmc for SFMC-aware PR reviews.

To enable it in your own repository:

1. Copy .github/agents/sfmc-reviewer.agent.md to your repo.
2. In your GitHub repo settings → Copilot → Cloud agent → MCP configuration, add:

{  "mcpServers": {    "sfmc": {      "type": "stdio",      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"],      "tools": ["*"]    }  }}

3. Assign the sfmc-reviewer agent to a pull request by mentioning it in a comment or via the @sfmc-reviewer agent in GitHub Copilot Chat.

GitHub Copilot (dedicated PR review)

Copy .github/copilot-instructions.md from this repo to your project. GitHub Copilot's dedicated PR review feature reads this file and applies the SFMC language rules when summarising your PRs.

GitLab Duo

1. Copy the content of ci-templates/gitlab-duo-review-instructions.md to .gitlab/duo-instructions.md in your repository.
2. GitLab Duo Code Review will apply these instructions on every merge request.

GitLab Duo's dedicated MR review does not support MCP directly. Use the CI lint job below for automated static analysis, and the Duo instructions for AI-assisted review comments.

CI linting (deterministic checks)

For deterministic, blocking CI validation, use the templates provided in ci-templates/. Each one runs two checks on every PR/MR: eslint-plugin-sfmc on changed files, and sfmc-review-diff (this package) on the unified diff — so a finding fails the job regardless of which AI reviewer (Bugbot, Copilot, Duo) is also enabled.

The ESLint job posts lint results as PR/MR comments; the sfmc-review-diff job exits non-zero on ERROR diagnostics by default (--fail-on warning|info to be stricter). See ci-templates/README.md for the difference between the two checks.

ESLint + @eslint/mcp

For AI assistants that don't support MCP but do support tool-calling, you can combine eslint-plugin-sfmc with the official @eslint/mcp server. Add it alongside mcp-server-sfmc:

{  "servers": {    "sfmc": {      "command": "npx",      "args": ["-y", "mcp-server-sfmc@latest"]    },    "eslint": {      "command": "npx",      "args": ["-y", "@eslint/mcp@latest"]    }  }}

Create an eslint.config.mjs in your project root:

import sfmc from 'eslint-plugin-sfmc';export default [...sfmc.configs.recommended];

The @eslint/mcp server exposes an eslint_lint tool that your AI can call to run the full ESLint rule set (including all AMPscript and SSJS rules from eslint-plugin-sfmc) on any file.

Architecture

mcp-server-sfmc    ├── sfmc-language-lsp   (language intelligence core)    │       ├── ampscript-data   (AMPscript function catalog)    │       ├── ssjs-data        (SSJS function catalog)    │       └── handlebars-data  (MCN Handlebars helper catalog)    ├── ampscript-data     (direct — conversion maps, mcnSince, handlebarsEquivalent)    ├── ssjs-data          (direct — Platform.Function ↔ AMPscript mapping)    └── handlebars-data    (direct — AMPscript ↔ Handlebars conversion)
vscode-sfmc-language (VS Code extension)    └── sfmc-language-lsp   (same core, bundled via esbuild)

Both the VS Code extension and the MCP server share exactly the same validation, completion, hover, and lookup logic through sfmc-language-lsp. This means the AI assistant sees the same errors and suggestions that the editor shows.

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.