Markaestro

@markaestro/mcp

A Model Context Protocol server that lets AI agents (Claude Code, Claude, Cursor, ChatGPT, Grok, Grok Bot, OpenClaw, Hermes, and any other MCP client) schedule, publish, and review Markaestro posts through the public API. Exact connection steps per client are on https://markaestro.com/developers/agents (?client=cursor, chatgpt, grok, grok-bot, openclaw, hermes, or headless opens that tab).

Every key covers either one brand or every brand in its workspace, chosen when the key is created or at the sign-in consent. A single-brand key works on that brand alone. An all-brands key sees every brand: list_products returns them all, and the agent passes productId to create_post, create_posts, list_posts, and create_evergreen_queue, and to the analytics tools to report on one brand instead of the whole workspace.

Tools

The Evergreen tools (preview_, create_, get_, list_, update_, activate_, pause_, and resume_evergreen_queue, list_evergreen_runs, get_evergreen_analytics) manage proven recurring content.

Agents manage social media and nothing else: account settings, billing, team, API keys, webhooks, and channel connections are not exposed, and no tool deletes a published post, takes one down from a platform, or archives an Evergreen queue. Those stay with the user in Markaestro. The same rule holds at the REST layer for a key issued through the agent sign-in: it cannot remove a published post, take one down from a platform, or manage webhooks (FORBIDDEN_AGENT_CONNECTION). A key a person mints in Settings keeps the full REST surface its scopes allow.

Also served: the markaestro://channel-rules resource and a schedule_post prompt that walks an agent through a safe scheduling flow.

Posting is draft-first. create_post without scheduledAt never publishes; publish_post is the only tool that publishes now, and its description tells the agent to confirm with the user first.

Two ways to connect

Hosted (nothing to install, nothing to paste). Markaestro serves the same tools over Streamable HTTP at https://markaestro.com/api/public/v1/mcp. Add it with no credentials and the first tool call opens the browser: sign in, pick the workspace and brand, click Allow. The client receives an API key bound to that brand and refreshes it on its own.

claude mcp add --transport http markaestro https://markaestro.com/api/public/v1/mcp

This is standard OAuth 2.1 (PKCE, dynamic client registration, discovery at /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server), so claude.ai, Cursor, and other MCP clients connect the same way. Connected agents are listed and revoked in Settings, API.

For headless or CI use, pass a workspace API key as a bearer header instead:

claude mcp add --transport http markaestro https://markaestro.com/api/public/v1/mcp \  --header "Authorization: Bearer mk_live_..."

{  "mcpServers": {    "markaestro": {      "type": "http",      "url": "https://markaestro.com/api/public/v1/mcp",      "headers": { "Authorization": "Bearer mk_live_..." }    }  }}

Send x-markaestro-read-only: 1 as an extra header to get only the reading tools.

Local package. Run the server on your machine over stdio, which also lets upload_media read local files:

Setup

1. Create an API key in Markaestro under Settings > API Access. Pick the brand, the scopes the agent needs (products.read, posts.read, posts.write, media.write, and posts.publish if it may publish), and prefer a test key while you evaluate.
2. Give the key to the server through MARKAESTRO_API_KEY.

Claude Code

claude mcp add markaestro -e MARKAESTRO_API_KEY=mk_live_... -- npx -y @markaestro/mcp

Claude Desktop, Cursor, and other JSON configs

{  "mcpServers": {    "markaestro": {      "command": "npx",      "args": ["-y", "@markaestro/mcp"],      "env": { "MARKAESTRO_API_KEY": "mk_live_..." }    }  }}

Environment

What the server does for you

• Sends an Idempotency-Key on every mutation, minted once per call, so a retried request replays instead of double-posting.
• Retries 429 and transient 5xx responses using the server's Retry-After.
• Runs the three-step direct media upload (session, PUT to storage, finalize) and never sends the API key to the storage URL.
• Returns API failures as tool errors with the stable error code, any per-channel issues, and a hint about what to change.

Claude Code plugin

The skill and the hosted server ship together as a plugin:

claude plugin marketplace add markaestro/markaestro-agentsclaude plugin install markaestro@markaestro

Deploying

Maintainers: the full release runbook (every surface, the version files, and the exact commands) is docs/operations/mcp-release.md in the Markaestro app repository. Summary:

Development

cd mcpnpm installnpm run buildMARKAESTRO_API_KEY=mk_test_... MARKAESTRO_BASE_URL=http://localhost:3000 npm run smoke            # local stdioMARKAESTRO_API_KEY=mk_test_... MARKAESTRO_BASE_URL=http://localhost:3000 npm run smoke -- --remote # hosted endpoint

Unit tests live in src/__tests__ and run with the repository's npm test. The smoke script starts the built server over stdio, lists tools, reads the brand, creates a draft, reads it back, and deletes it. Nothing is published.