Eds MCP Server

Let an AI agent run — and *improve* — your Adobe Edge Delivery site.

41 tools. No extra dependencies beyond the MCP SDK. Works with any EDS site. The first MCP server purpose-built for Edge Delivery Services.

Read your content → audit it → fix what's wrong → publish → undo any of it. One page or the whole site, in a single reversible operation. Preview before every write; undo after.

---

⚡ Do it in three lines

claude mcp add eds -e EDS_OWNER=your-org -e EDS_REPO=your-site -- npx @focusgts/eds-mcp-server

Then just ask your agent:

*"Audit the whole site and show me what's hurting SEO."*

*"Fix the meta description on every page that's missing one — preview first, then publish."*

*"Actually, undo that whole batch."*

*"Preview and publish the homepage."*

That's it — no local AEM, no scripts, no glue code. Every write is previewable and reversible.

The loop that makes it different

flowchart LR

A["📖 Read authored content (DA)"] --> B["🔍 Audit SEO · a11y · perf · freshness"] B --> C["🛠️ Fix one page or the whole site"] C --> D["🚀 Publish preview + live"] D --> E["↩️ Undo one call reverts it all"] E -. "re-audit to zero" .-> B

It doesn't just *drive* your site — it improves it, safely. Point it at an EDS site and an agent can find what's wrong and repair it, one page or the whole site in a single batch that a single eds_da_rollback reverts. No other MCP server — including Adobe's own — does this end-to-end.

---

🧠 How it works

flowchart LR

A["AI agent (Claude Code · Cursor · Copilot)"] -- MCP / stdio --> B["eds-mcp-server 41 tools"] B --> C["Admin API admin.hlx.page"] B --> D["Content API *.aem.live"] B --> E["RUM / OpTel Core Web Vitals"] B --> G["Document Authoring admin.da.live"] C --> F["Your EDS site"] D --> F E --> F G --> F

The agent calls tools; the server talks to the live EDS infrastructure. Read-only tools (content, sitemap, metadata) need no credentials at all.

---

🔑 One-click sign-in

No more pasting a fresh admin token every day:

sequenceDiagram

participant You participant CLI as eds-mcp-server login participant Adobe as admin.hlx.page You->>CLI: npx @focusgts/eds-mcp-server login CLI->>Adobe: open browser (client_id=aem-cli) You->>Adobe: sign in & approve Adobe-->>CLI: siteToken → localhost callback CLI-->>You: cached ~/.aem/auth-token.json (reused automatically)

Use Chrome or Firefox — Safari blocks the local callback (same as Adobe's AEM CLI). EDS_API_KEY works as the CI / fallback path.

---

🛠️ The 41 tools

Edge Delivery Services — publish, content, analytics

Publishing

• eds_preview_page
• eds_publish_page
• eds_unpublish_page
• eds_preview_and_publish
• eds_get_status
• eds_purge_cache
• eds_bulk_preview
• eds_bulk_publish
• eds_get_job_status

Content

• eds_get_page
• eds_list_pages
• eds_search_pages
• eds_get_metadata
• eds_get_sitemap
• eds_get_redirects

Analytics & config

• eds_get_cwv
• eds_get_404s
• eds_get_experiments
• eds_get_config
• eds_get_logs
• eds_get_api_keys

Document Authoring (DA) — the authored *source*, not the rendered output

Nine tools reach a site's Document Authoring source directly (admin.da.live), the source of truth behind an EDS site. Requires EDS_DA_TOKEN.

Read

• eds_da_list_sources
• eds_da_get_source
• eds_da_get_versions

Write

• eds_da_put_source
• eds_da_delete_source
• eds_da_copy_source
• eds_da_move_source

Bulk ("clone") + safe writes

• eds_da_export
• eds_da_push
• eds_da_rollback

`eds_da_export` / `eds_da_push` bring the efficiency of aem content clone to agents: export a whole DA subtree in one call, operate on it, and push the batch back in one call — no local checkout, no aem-cli. Same model, network-native.

Safe by default. eds_da_push takes dryRun: true to preview exactly what a bulk edit would do (create / update / unchanged, with line-diff counts) without writing a thing, and withUndo: true to make the write reversible — it returns an undo object you hand to eds_da_rollback to restore prior content and remove any docs the push created. Preview before writing, undo after: the difference between an impressive demo and something you'd point at a production site.

EDS_DA_TOKEN is an Adobe IMS access token for Document Authoring — grab it from an authenticated da.live session (the IMS access_token). Document paths assume .html when no extension is given (index → index.html).

Content audit — find what's wrong, before you fix it

• eds_audit_page
• eds_audit_site
• eds_audit_report
• eds_audit_snapshot
• eds_audit_trend
• eds_audit_monitor

It tells you what's wrong. eds_audit_site sweeps the whole site (or a subtree) and returns a prioritized list of issues across SEO (missing titles/descriptions, no H1, blocked from indexing), accessibility (images without alt text, missing landmarks, unlabeled form inputs), freshness (pages not updated in over a year), sitemap coverage, and — with a domain — performance (Core Web Vitals) and 404s from Adobe's own real-user data. eds_audit_page does the same for one page. Read-only and safe to run anytime.

`eds_audit_report` turns that audit into a beautiful, client-ready HTML report — a Focus GTS Navigator letterhead, an executive summary, per-dimension health scores, a prioritized issue list with each suggested fix, and a Save-as-PDF button (uses your browser's own print — no dependency). Self-contained (no external assets), ready to open, host, or send to a stakeholder. Pass an optional brand (agency name, logo, accent, "prepared for" client) to white-label the letterhead.

Track it over time. eds_audit_snapshot records each audit's scores to a history sheet in your site's own content (private by default) and tells you the change since last time — *"89, ▲7 since last week."* `eds_audit_trend` turns that history into a shareable HTML sparkline of your score over time plus per-dimension movement. One snapshot is a mirror; the trend is the story.

Watch it on autopilot. eds_audit_monitor audits, diffs against the last snapshot, and reports a status — ok / degraded / broken — and, when you give it a webhook, pings Slack/Discord the moment health breaks (a new critical, or a dimension fallen to poor). The server does the check + alert; you supply the schedule — a copy-paste [scheduled GitHub Action](examples/monitor.yml) or your agent runtime. Webhook is https-only and the payload carries no secrets.

Safe fixes — repair what the audit finds

• eds_fix_metadata
• eds_bulk_fix_metadata
• eds_fix_redirect
• eds_fix_audit

It fixes what it finds — reversibly. eds_fix_metadata repairs a page's title, meta description and Open Graph image by editing its Document Authoring source, routed through the same dry-run + undo path as the write tools. The agent supplies the content (e.g. writes a fitting description); the tool writes it *correctly and idempotently* (merges into the page's Metadata block, never duplicates it). Pass publish: true to preview + publish so the change goes live.

`eds_bulk_fix_metadata` does it across a whole site in one reversible operation — pass a list of { path, metadata }, and it writes every changed page in a single batch that returns one undo reverting all of it. The full loop: `eds_audit_site` → fix the batch → publish → re-audit to zero — with a single undo if anything looks off.

`eds_fix_redirect` closes the 404 loop: eds_audit_site surfaces the broken links from real-user data, and this adds the 301 redirect rules (to the site's redirects sheet) that fix them — one rule or many, idempotent, dry-run + undo. So the audit now has a fix for *every* major finding.

`eds_fix_audit` is the "fix it" button in agent form: after an audit, apply its fixable findings — metadata and redirects together — in one reversible batch. Findings the report marks ✦ Fixable carry a machine-readable fix; you supply the values (the tool never invents copy), and every change is pushed at once so a single eds_da_rollback undoes all of it. dryRun previews the whole plan; publish: true makes it live.

---

🔌 Add it to your tool

Claude Code — one command

claude mcp add eds -e EDS_OWNER=your-org -e EDS_REPO=your-site -- npx @focusgts/eds-mcp-server

Cursor — .cursor/mcp.json

{

"mcpServers": { "eds": { "command": "npx", "args": ["@focusgts/eds-mcp-server"], "env": { "EDS_OWNER": "your-org", "EDS_REPO": "your-site" } } } }

</details>

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

{

"servers": { "eds": { "command": "npx", "args": ["@focusgts/eds-mcp-server"], "env": { "EDS_OWNER": "your-org", "EDS_REPO": "your-site" } } } }

</details>

---

⚙️ Configuration

Read-only tools (content, sitemap, metadata) need no keys. Write tools (preview, publish, cache) need an admin token. Analytics tools need EDS_DOMAIN_KEY. DA source tools need EDS_DA_TOKEN.

---

🔐 Authentication

Admin operations require an EDS Admin token. Two ways to provide one.

Browser sign-in (recommended for interactive use)

EDS_OWNER=your-org EDS_REPO=your-site npx @focusgts/eds-mcp-server login

Opens your browser to Adobe's admin.hlx.page login (the same flow as the AEM CLI). The admin site token caches at ~/.aem/auth-token.json (mode 0600, ~24h) and is reused automatically. **Use Chrome or Firefox — Safari blocks t