Postmcp Mcp Server

PostMCP AI Model Context Protocol (MCP) Server

[npm version] [License: MIT] [MCP Compatible]

Official PostMCP AI Model Context Protocol (MCP) Server. Connect your social media publishing pipelines directly into AI assistants, desktop applications, IDE workflows, and web environments like Claude Desktop, Claude.ai, Cursor, and ChatGPT Custom GPTs.

Supported platforms include LinkedIn, X (Twitter), Facebook, Instagram, Threads, Bluesky, and YouTube Shorts.

🚀 Features & Capabilities

• 🤖 20 Built-in Tools: Workspaces, connected accounts and their token health, the post queue, pre-flight checks, create/schedule/reschedule/publish/retry/delete, per-post and per-profile analytics, media uploads, image generation, and batching.
• 🖼️ Carousels & galleries: Pass mediaUrls to publish a carousel on Instagram and Threads, a multi-photo post on Facebook and LinkedIn, or a four-image gallery on X and Bluesky - one call, every network's ceiling checked up front.
• ⚡ Dual Transport Modes: Native Stdio mode (for local desktop apps & IDEs) and Streamable HTTP mode (for web services, Claude.ai, and remote connectors).
• 🔑 Flexible Authentication: Auto-detects API key from environment variables (POSTMCPAI_API_KEY), URL query parameters (?apikey=YOUR_KEY), or HTTP authorization headers (x-api-key, Bearer token).
• 🗂️ Multi-Workspace Aware: The API key carries its own workspace, so a bare key is enough. To act on another one, every tool takes an optional workspaceId, also settable per connection (?projectId=..., x-project-id) or per process (POSTMCPAI_PROJECT_ID).
• 🤖 ChatGPT Actions Compatible: Includes built-in OpenAPI 3.0 specification generator (/openapi.json) and REST endpoints (/api/tools/:name) for ChatGPT Custom GPT integration.
• 🔒 OAuth with PKCE and RFC 9728: Per-user consent, persistent encrypted grants, rotating tokens and DCR for ChatGPT/Codex. Requires the production settings in OAUTH.md; API-key clients remain supported.
• 🧩 Installable plugin: Account onboarding, publishing, scheduling and analytics workflows. See PLUGIN.md and run npm run package:plugin.

📁 Repository Architecture

mcp-server/├── bin/│   └── cli.js            # Executable CLI entry point (Stdio / HTTP mode runner)├── src/│   ├── config.js         # Centralized configuration & environment loader│   ├── client.js         # Backend API client, API key & workspace extraction│   ├── platforms.js      # Platform limits, credit pricing & post cost helper│   ├── tools/│   │   ├── definitions.js# MCP tool JSON schemas & parameter specifications│   │   ├── handlers.js   # MCP tool execution handlers│   │   └── index.js      # Tool definitions aggregator│   ├── server.js         # MCP Server instance factory│   ├── routes/│   │   ├── oauth.js      # OAuth 2.0 & RFC 9728 discovery endpoints│   │   ├── openapi.js    # OpenAPI 3.0 schema & ChatGPT REST endpoints│   │   ├── mcpHttp.js    # MCP Streamable HTTP transport (/mcp)│   │   └── health.js     # Health check & system metadata endpoints│   ├── app.js            # Express application factory│   └── index.js          # Main library entry point├── index.js              # Executable wrapper script├── package.json└── README.md

⚙️ Environment Configuration

🛠️ MCP Tools Reference

Every tool below also accepts an optional workspaceId (from list_workspaces) to act on a specific workspace.

Reading

Writing

Media uploads

For a local file, the MCP client needs an HTTP upload capability:

1. Call create_media_upload_url with fileName, contentType (MIME type), and fileSize (bytes). For example: { "fileName": "launch.png", "contentType": "image/png", "fileSize": 245760 }.
2. PUT the raw file bytes to the returned uploadUrl with the returned headers. Do not send your PostMCP API key to storage. The URL expires after expiresIn seconds (normally 900); request another if it expires. The file is not uploaded merely by creating the URL.
3. After the PUT succeeds, call complete_media_upload with the returned publicUrl as url, the returned key, and the same file metadata. Use the same workspaceId on both tools. This registers the file; it does not transfer or verify the bytes.
4. Use the returned mediaUrl in create_post or update_post, or collect several URLs in mediaUrls for a carousel.

For an existing public file, call import_media_from_url with { "url": "https://example.com/launch.png" }. The backend downloads it and returns the stored mediaUrl in one call. The URL must serve the file directly without authentication; private-network URLs are rejected by the backend.

Storage supports PNG, JPEG, GIF, WebP, AVIF, HEIC, MP4, MOV, WebM and M4V, with a 25 MiB image ceiling and a 512 MiB video ceiling. Individual social platforms have stricter publishing requirements; use preflight_post before publishing. File bytes and base64 never travel through the MCP JSON request. If library registration fails after storage succeeds, the result includes librarySaved: false and a warning while preserving the usable mediaUrl.

Batching

{  "calls": [    { "id": "img", "tool": "generate_image", "arguments": { "prompt": "launch banner" } },    {      "tool": "create_post",      "arguments": {        "content": "We shipped it 🚀",        "targetAccounts": [          { "platform": "linkedin", "profileId": "lin_7741903" },          { "platform": "twitter", "profileId": "tw_1293847", "content": "We shipped it 🚀" }        ],        "scheduleDate": "2026-09-01",        "scheduleTime": "10:00",        "timezone": "Asia/Kolkata"      }    }  ],  "stopOnError": true}

The reply carries one entry per call — { id, tool, ok, result } or { id, tool, ok: false, error } — plus counts and, when a failure stopped the batch, the calls that were skipped.

Carousels

mediaUrls is the ordered attachment set. One URL is an ordinary media post; two or more publish as a multi-media post on every network but YouTube, with the first URL as the cover:

{  "content": "Five things we learned shipping v2 👉",  "targetAccounts": [    { "platform": "instagram", "profileId": "17841400000000" },    { "platform": "threads", "profileId": "9988776655" },    { "platform": "twitter", "profileId": "tw_1293847",      "mediaUrls": ["https://cdn.example.com/v2/1.png", "https://cdn.example.com/v2/2.png", "https://cdn.example.com/v2/3.png", "https://cdn.example.com/v2/4.png"] }  ],  "mediaUrls": [    "https://cdn.example.com/v2/1.png",    "https://cdn.example.com/v2/2.png",    "https://cdn.example.com/v2/3.png",    "https://cdn.example.com/v2/4.png",    "https://cdn.example.com/v2/5.png"  ],  "scheduleDate": "2026-09-01",  "scheduleTime": "10:00",  "timezone": "Asia/Kolkata"}

• Only Instagram and Threads mix video into a carousel; everywhere else a set of several must be images only, and a video goes out on its own.
• create_post refuses a set a target will not take before any credits are spent, naming the profile and the rule. preflight_post with the same mediaUrls reports the same thing plus a mediaSetLimits map, so check first when one carousel goes to several networks.
• A profile can carry its own mediaUrls on its targetAccounts entry (or in variants as { "twitter": { "mediaUrls": [...] } }), replacing the shared set - the way to give X and Bluesky a four-slide cut of a longer carousel.
• Every slide is copied into PostMCP's own storage at write time, like a single attachment, so a host that expires the links later does not break the scheduled post. One slide failing at publish time fails that profile's post rather than publishing a shorter carousel; publish_post_now retries it.
• update_post with mediaUrls replaces the whole set (add, remove or reorder slides); an empty array removes all media. Every post returned by list_posts / get_post carries mediaUrls alongside mediaUrl.

Notes for clients

• Target profiles, not platforms. targetAccounts sends only to the profiles named; platforms fans out to every connected profile on each platform.
• One post per profile. create_post stores a separate post per targeted profile, so each can be edited, retried or cancelled on its own. Give per-profile copy through targetAccounts[].content or the variants map.
• Always pass timezone when a wall-clock time matters. The backend defaults to UTC, so a 9:00 IST post scheduled without a zone goes out at 14:30 IST.
• Credits are charged per profile delivered to (X/Twitter costs 5, others 1), plus a one-off 50-credit surcharge when the copy contains a link. preflight_post reports this before you commit.
• Analytics are read on request, never in the background. list_posts and get_connected_accounts carry what the last reading stored, free. get_post_analytics (one post, every network it went to) and get_profile_analytics (one connected profile: followers, posts, views) with refresh: true read the network now for 1 credit each; without it, the stored reading is free. engagements is likes + comments + shares on every network, so it compares across platforms. A null means the network does not report that metric (Bluesky has no views), an error naming reconnect means the account predates the insights permission and its owner must reconnect it, and unavailable: true means the network never answers for that kind of post (LinkedIn personal profiles).

💻 Client Integration Guides

1. Claude Desktop App (Stdio Mode)

Add the configuration below to your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{  "mcpServers": {    "postmcpai": {      "command": "npx",      "args": ["-y", "@postmcpai/server"],      "env": {        "POSTMCPAI_API_KEY": "pmcp_sec_your_secret_api_key_here"      }    }  }}

2. Cursor IDE

1. Open Cursor Settings -> Features -> MCP.
2. Click + Add New MCP Server.
3. Fill in the details:
   • Name: postmcpai
   • Type: command
   • Command: npx -y @postmcpai/server
4. Under Environment Variables, add:
   • POSTMCPAI_API_KEY = pmcp_sec_your_secret_api_key_here
5. Click Save.

3. Claude.ai & Remote Web Connectors (Streamable HTTP / SSE Mode)

Host this server on any cloud service (Render, Railway, Fly.io, Vercel) or tunnel your local machine using ngrok.

Launching in HTTP Mode:

export PORT=3000# Each remote caller supplies its own API key or OAuth token.
npm run start:sse

Connecting to Claude.ai:

1. Provide your public MCP URL with your API key attached: https://your-hosted-domain.com/mcp?apikey=pmcp_sec_your_secret_api_key_here
2. Claude.ai will discover tool capabilities via /mcp and authenticate seamlessly.
3. That URL is all you need: the key is bound to the workspace it was issued from, so tools act on that workspace without being told. To point the same key at a different workspace, append &projectId=YOUR_WORKSPACE_ID (or send an x-project-id header); individual tool calls can still override either with workspaceId.

4. ChatGPT Custom GPTs (REST Actions)

1. When configuring a Custom GPT Action, specify your server URL (e.g. https://your-hosted-domain.com).
2. Import the OpenAPI schema directly from: https://your-hosted-domain.com/openapi.json
3. Set Authentication to API Key (Bearer Authorization or custom x-api-key). This is a GPT Actions integration, not a per-customer OAuth plugin. For the plugin, use OAUTH.md.

5. Programmatic Node.js Library Usage

You can also use @postmcpai/server as a library in your own Node.js backends:

import { createServer, createExpressApp, makeBackendRequest } from "@postmcpai/server";
// Create a standalone MCP Server instanceconst mcpServer = createServer(() => process.env.POSTMCPAI_API_KEY);
// Or create an Express app with all remote routes attachedconst app = createExpressApp();app.listen(3000);

🧪 Local Testing & Development

# Clone the repositorygit clone https://github.com/postmcp/postmcp-mcp-server.gitcd postmcp-mcp-server
# Install dependenciesnpm install
# Start in Stdio Modenpm start
# Start in HTTP Mode with hot reloadnpm run dev

📄 License

Distributed under the MIT License. Copyright © 2026 PostMCP AI.