Layers for agents

Run the department over MCP

Point Claude, Codex, or any MCP client at Layers and it gets real tools: onboard an account, generate and approve content, publish on a schedule, and read what it earned. You never have to open the web app.

One install

Add the server, get the tool set

Node 20+ and an integration key. There is nothing to install globally and no plugin to approve: npxfetches the server on first run.

Claude Code
$ claude mcp add layers \
--env LAYERS_API_KEY=lp_YOUR_KEY \
-- npx -y @layers/mcp-server@latest
✓ Layers tools available in your next session

Claude Desktop

Add it to claude_desktop_config.json

OpenAI Agents SDK

Launch it with MCPServerStdio

Any MCP client

Same command and args over stdio

Node 20+ · stdio · the key is the whole permission boundary

Onboard without ever opening the app

Point the server at a URL and it starts an account from nothing: your brand read off the site, questions asked, a starter plan generated. You get a preview link to look at what it made, and a claim link when you want to keep it. No signup form first.

Keyless startNo account yet

onboard https://myapp.com

✓ brand read · plan drafted

preview /p/8f3a…

claim /claim?token=…

When you need to look, it hands you a link

Some things are worth seeing rather than reading back as JSON. Ask for a dashboard or a piece of content and the server mints a link to a rendered view of exactly that. You open it, look, and go back to the thread.

Rendered viewOpens in the browser
Contentthe post, as it will publish
Metricsthe dashboard, live
Planthe week, as scheduled

Nothing the key can't already do

The key is the entire permission boundary. Its organization, scopes, and limits apply exactly as they do for a direct call, and the server adds no privileges of its own. Run it read-only and the mutating tools aren't registered at all.

Key modesSet at launch
lp_live_…Read and writeFull surface
--read-onlyWrites not registeredReads only
lp_test_…Sandbox · nothing shipsFull surface

The surface

Every tool is a real endpoint

Each tool maps one-to-one to a documented API route. The long-running ones return a job you poll, so an agent can start a render and carry on.

Creative

Generate slideshows, UGC remixes, and video remixes. Read hooks and keywords, upload your own footage, approve or reject what comes back.

Distribution

Publish or schedule to connected accounts, reschedule and cancel posts, read the queue, and set the engagement config.

Core

Projects and identity: create, read, update, archive. Check credits and read the credit ledger.

Measurement

Unified metrics, top performers, ads content, and the recommendations the loops produce.

Framework

The audit log: every action taken through the key, readable back.

Full reference

Every tool →

Give your agent the department

One command, a key, and the whole surface is available in your next session.