@upload-post/mcp
Official Model Context Protocol (MCP) server for Upload-Post.
Lets any MCP-compatible AI agent (ChatGPT, Claude Desktop, Claude Code, Cursor, …) publish, schedule, analyze and manage social media across TikTok, Instagram, YouTube, LinkedIn, Facebook, Pinterest, Threads, Reddit, Bluesky, X, Google Business, Discord, Telegram and more with a single API key.
Built on top of the official
upload-postSDK and the public Upload-Post REST API.
Use it in ChatGPT (no setup)
Upload-Post is a plugin reviewed by OpenAI, listed in ChatGPT's Plugins (formerly Apps). Open Upload-Post in ChatGPT or search for Upload-Post in Plugins, click Install, then Connect, and sign in with OAuth. There is no MCP URL to paste, and you get the Upload Studio widget for publishing video files, photos and carousels straight from your computer. OpenAI dots use the same plugin: select it in your dot's Customize → Plugins and review its permissions.
Prefer to wire it up by hand? In ChatGPT on the web, open Plugins, click +, choose Add custom MCP server, point it at https://mcp.upload-post.com/mcp and set authentication to OAuth. Custom MCP servers get write actions on Business, Enterprise and Edu workspaces.
Use it in Claude
In claude.ai or Claude Desktop, open Customize → Connectors, click + Add, then Add custom connector, paste https://mcp.upload-post.com/mcp, click Continue and sign in with OAuth. Custom connectors work on every Claude plan (the Free plan allows one). In Claude Code: claude mcp add --transport http upload-post https://mcp.upload-post.com/mcp.
Use it in Perplexity Computer
Open Plugins in the sidebar (or Customize → Connectors), then Add custom → Connector → Add MCP connector → Remote. Use https://mcp.upload-post.com/mcp with the Streamable HTTP transport and OAuth (or your Upload-Post API key), click Add, then click the Upload-Post card to sign in.
Use it in Gemini CLI
Install the extension and sign in with OAuth the first time Gemini calls a tool:
gemini extensions install https://github.com/Upload-Post/upload-post-mcp
Two ways to run it yourself
A) Local stdio (single-user) — simplest
The server runs on your machine, spawned by the MCP client. Add to ~/.claude/mcp.json (or Cursor settings, etc.):
{
"mcpServers": {
"upload-post": {
"command": "npx",
"args": ["-y", "@upload-post/mcp"],
"env": { "UPLOAD_POST_API_KEY": "YOUR_API_KEY" }
}
}
}
Get your API key at https://app.upload-post.com → API Keys. Restart the client — you should see 58 upload-post tools.
B) Hosted HTTP (multi-tenant) — share one server with many users
Run the server on any Docker-capable host (Fly, Railway, Cloud Run, your own box…) and let each user connect with their own Upload-Post API key. The server stores nothing per user.
{
"mcpServers": {
"upload-post": {
"url": "https://mcp.your-domain.com/mcp",
"headers": {
"Authorization": "ApiKey YOUR_OWN_UPLOAD_POST_API_KEY"
}
}
}
}
Authorization: Bearer <key> is also accepted, for clients that only allow Bearer.
What can the agent do?
The server exposes Upload-Post API tools plus one ChatGPT App UI launcher.
Async uploads return a request_id. The agent should poll get_status until success: true.
One tool per question, not per network
The Upload-Post API has no endpoint per social network: it has an endpoint per question, and a platform parameter saying who is being asked. get_post_comments, get_post_analytics, get_audience, get_suggestions and comment_action all work that way, so an agent learns one shape and reuses it for every network. Only the four tiktok_* tools are network-specific, because what they return (the Commercial Music Library, TikTok places, TikTok's per-account publishing settings) exists only on TikTok.
get_audience— who follows the profile, where they are, when they are online, what they tap. Alsobenchmark_categories, and the niche averages to compare against whenbenchmarkCategoryis set. The server clamps the window to at most 60 days ending before today, so a wider range is trimmed rather than rejected, andrangein the response says which window was used.get_suggestions— hashtags (withview_count) or related keyword searches, told apart bytype, not by a different tool.get_post_comments— top-level comments on a post, or, withcommentId, the replies under one of them.hide_comment,like_comment,pin_comment,edit_comment,hold_comment,approve_comment,ignore_comment,set_post_comments— one tool per moderation operation.comment_actiondoes the same through a singleactionargument and stays available everywhere except ChatGPT, which only gets the one-operation tools (the same goes formanage_autodms).comment_action— moderate a comment on TikTok (hide / like / pin), Facebook (hide / like / edit), Instagram (hide, or enable / disable comments on a post), YouTube (hide / hold) and Threads (hide / approve / ignore). Each value carries its own inverse, so nothing is permanent.get_post_analytics— per-post metrics.post_metricsis whatever the platform reports, so its shape varies: on TikTok it addsretention,impression_sources,audience_types,new_followers,reachand the watch times (average_time_watched,total_time_watched,full_video_watched_rate).
Errors are shared too: platform_not_supported (400, with the list of the networks that can answer), invalid_parameter (400), tiktok_reconnect_required (400), reauth_required (409) and 502 when the upstream network fails.
TikTok capabilities
get_post_comments, create_comment, delete_comment and comment_action accept platform: "tiktok", and firstComment works on TikTok like on every other network.
What a TikTok account can do depends on how it is connected. list_users returns a capabilities array on each TikTok account — music, location, cover_image, cover_timestamp, draft, video_privacy, photo_privacy, profile_analytics, comments, trend_search — and each tool's description names the one it needs. They are granted when the user connects TikTok, so an account connected before a capability existed has to reconnect before the matching tools answer; that is what a tiktok_reconnect_required error means.
get_media and get_cached_post_analytics are cursor-paginated: feed the response's next_cursor back as cursor until has_more is false. LinkedIn, Discord and Telegram do not support media cursors and accept limit only. Prefer get_cached_post_analytics over get_post_analytics when scanning many posts — it replays previously fetched results and so avoids the live analytics rate limit of 100 requests / 5 minutes. Only contains posts previously fetched through a live per-post endpoint; there is no background refresh, so captured_at is the last time that post was read live.
FFmpeg jobs accept one public URL through input_url or multiple URLs through files. They run on Upload-Post's encoding servers, each in its own working directory. full_command is a single ffmpeg invocation with {input} (or {input0}, {input1}…) and {output} placeholders, run without a shell: it must start with ffmpeg, and shell metacharacters, .., file: and absolute paths outside the job directory are rejected. Poll get_ffmpeg_job until completion, then call download_ffmpeg_result; it returns the result URL without streaming the processed binary through MCP. Nothing is posted until you pass that URL to an upload tool.
Confirmation before acting
Tools that publish or reach other people as the user run in two steps. Called without confirmed: true, they return status: "confirmation_required" with a preview of exactly what would go out (text, media, platforms or recipients, buttons, keywords, timing) and do nothing; the model shows it to the user and repeats the call with confirmed: true once they approve. This covers upload_video, upload_photos, upload_text, upload_document, send_dm, reply_to_comment, public_reply_to_comment, create_comment, reply_to_google_business_review, edit_comment, unpublish_post, start_autodm and manage_autodms with action: "start", plus comment_action / hide_comment when they edit a comment or ban its author. delete_user deletes only when confirm_username repeats the profile name. ChatGPT, which asks the user before every write tool, keeps the single-step tools.
Reddit publishing is switched off while the Reddit app is unavailable: reddit is not offered as a platform and asking for it fails with reddit_unavailable. Self-hosters can set UPLOAD_POST_REDDIT_ENABLED=true once it is back.
ChatGPT video upload UI
open_upload_studio renders a ChatGPT Apps component for file-based publishing: one video, or one or more images (up to 10, for a carousel; pass mediaType: "photos"). For each file the widget creates a short-lived Upload-Post/R2 staging upload, PUTs the file directly to R2 and completes the upload, one file at a time, then calls upload_video or upload_photos with the returned temporary media URLs in order. For photos it can instead hand the staged URLs back to ChatGPT ("Send to ChatGPT"), which then calls upload_photos itself, e.g. with a different caption per platform. upload_video and upload_photos refuse /mnt/data and sandbox attachment paths and point the model at the Studio.
The widget speaks the ChatGPT Apps SDK bridge, so it is only advertised to ChatGPT (detected from clientInfo at initialize; override the match with UPLOAD_POST_STUDIO_CLIENTS=<regex>). Every other host (claude.ai, Claude Desktop, Claude Code, Cursor, …) does not see the tool or its ui:// resource; instead create_media_upload / complete_media_upload are exposed to the model with step-by-step guidance, and upload_video tells the assistant to ask for a public URL or send the user to the dashboard when the client cannot PUT the file itself.
The staging object is deleted after 24 hours whether it is used or not. Scheduled/queued posts remain safe because upload_video copies the temporary URL into the existing durable scheduler storage before execution.
Claude and other MCP clients can use the same flow without the ChatGPT UI: call create_media_upload, PUT the file to upload_url, call complete_media_upload, then pass media_url to upload_video.
Set UPLOAD_POST_R2_CONNECT_DOMAINS on the MCP host to the comma-separated origins used by the backend's R2 signed URLs when they differ from the defaults (for example https://<account>.r2.cloudflarestorage.com,https://<bucket>.<account>.r2.cloudflarestorage.com) so the ChatGPT component CSP allows the browser PUT.
The R2 bucket CORS policy must allow browser uploads. A restrictive policy can include your actual widget origin; for fastest validation, use:
[
{
"AllowedOrigins": ["*"],
"AllowedMethods": ["PUT", "GET", "HEAD"],
"AllowedHeaders": ["*"],
"ExposeHeaders": ["ETag"],
"MaxAgeSeconds": 3600
}
]
Local / development
git clone https://github.com/Upload-Post/upload-post-mcp.git
cd upload-post-mcp
npm install
npm run build
# stdio (default — used by Claude Desktop, Cursor)
UPLOAD_POST_API_KEY=... node dist/index.js
# HTTP streamable (for hosted deployments)
UPLOAD_POST_API_KEY=... node dist/index.js --http --port 8080
Inspect the live tool surface with the official inspector:
npx @modelcontextprotocol/inspector node dist/index.js
Configuration
CLI flags:
--http— start the streamable HTTP transport instead of stdio--port <n>— port for HTTP mode
HTTP endpoints:
POST /mcp— JSON-RPC over MCP streamable HTTP. RequiresAuthorization: ApiKey <key>(orBearer <key>) on every request. The key is the user's own Upload-Post API key; the server uses it only for that session and stores nothing.GET /healthz— liveness probe, always open. Returns{"ok":true}.
Auth model in --http mode is the same pattern Resend, Tavily, Brave Search and other API-key-native services use for their hosted MCPs: the upstream key is the auth.
Deploy with Docker
The repo ships with a multi-stage Dockerfile and a .dockerignore. On any Docker-capable PaaS (Fly.io, Railway, Render, Cloud Run, fly machines, your own box…):
- Point the PaaS at this repo and select Dockerfile as the build pack.
- Port: 8080 (matches
EXPOSE 8080). - Environment variables: none are required. Optionally set
UPLOAD_POST_BASE_URLif you point at staging. - Health check path:
/healthz(HTTP, port 8080). - Domain: attach a domain, e.g.
mcp.your-domain.com, and provision TLS (most PaaS do this automatically via Let's Encrypt).
Deploy. The server is now ready for any number of users. Each user adds the endpoint to their MCP client config with their own Upload-Post API key:
{
"mcpServers": {
"upload-post": {
"url": "https://mcp.your-domain.com/mcp",
"headers": {
"Authorization": "ApiKey USER_OWN_UPLOAD_POST_API_KEY"
}
}
}
}
Without an
Authorizationheader the server returns401. The header is the only credential — invalid Upload-Post keys will surface as upstream errors on the first tool call.
Local test of the production image:
docker build -t upload-post-mcp .
docker run --rm -p 8080:8080 upload-post-mcp
curl http://localhost:8080/healthz # → {"ok":true}
curl -i -X POST http://localhost:8080/mcp \
-H "content-type: application/json" \
-d '{}' # → 401 (no Authorization)
Tips for prompting the agent
- Prefer public URLs over local paths when uploading — local paths only work if the MCP server runs on the user's machine.
- In ChatGPT Apps, prefer
open_upload_studiofor user-selected video and image files. It avoids local-path handoff issues by uploading to short-lived Upload-Post/R2 staging, then passing temporary media URLs toupload_video/upload_photos. - In any other client with a local file: if the client can run HTTP requests (Claude Code, Cursor, a script), stage it with
create_media_upload→ PUT the bytes toupload_url→complete_media_upload, then pass the returnedmedia_urltoupload_video. Hosted chats without that ability (claude.ai) need a public HTTPS URL or the dashboard at https://app.upload-post.com. - To send video bytes directly (a client that holds the file rather than a URL), pass
videoBase64toupload_videoinstead ofvideoPathOrUrl. The server writes it to a temp file, uploads, then deletes it. Inline bytes are capped atUPLOAD_POST_MAX_INLINE_MB(default 100 MB), and in--httpmode the whole request atUPLOAD_POST_MCP_MAX_BODY_MB(default 10 MB, roughly 7 MB of video once base64-encoded) — for larger videos use a public URL. - Always create the profile first (
create_user) and connect socials in the Upload-Post dashboard before publishing. - For scheduled posts, pass ISO 8601 dates with timezone, e.g.
"2026-12-25T10:00:00Z"+"timezone": "Europe/Madrid".
Privacy & data handling
This server is a stateless proxy to the Upload-Post API. Per request, the only data it processes is the user's API key (or OAuth access token resolved to one) and the arguments of the tool call being executed. No user data is persisted by the MCP container itself.
- What we receive per request: the
Authorizationheader, the MCP tool name + arguments, and any media URLs/paths the agent passes. - What we forward: the tool arguments to the Upload-Post API on behalf of the authenticated user.
- What we store: nothing per-user. OAuth tokens are stored upstream in the Upload-Post backend, hashed (SHA-256), so a breach of token storage cannot impersonate users.
- What we log: HTTP method, path, status code, and an opaque request ID. No tool arguments, no API keys, no tokens.
Full Upload-Post privacy policy (data collection, retention, third-party sharing, contact, GDPR/CCPA): https://upload-post.com/privacy
To revoke a connector's access at any time, open Connected Apps in app.upload-post.com.
Security
- All traffic is TLS-terminated at the edge (HTTPS only).
/mcprequires a validAuthorizationheader on every request; OAuth access tokens are short-lived (1 h access + 90 d refresh with rotation per RFC 6749 §10.4).- The server validates the
Originheader against an allow-list (claude.ai,claude.com,chatgpt.com,chat.openai.com,app.upload-post.com,localhost) to mitigate DNS-rebinding attacks from browser-based clients. Extend withOAUTH_EXTRA_ALLOWED_ORIGINS(comma-separated) when self-hosting behind a custom dashboard. - OAuth metadata is served both at
/.well-known/oauth-protected-resourceand at the path-inserted form/.well-known/oauth-protected-resource/mcp(RFC 9728 §3.1), and likewise foroauth-authorization-server. - In
--httpmode the server writes one stderr line per failed request and per OAuth/well-known request (method, path, status, auth scheme, user agent — never query strings or credentials). SetUPLOAD_POST_MCP_ACCESS_LOG=allto log every request oroffto disable it. - If ChatGPT shows
redirect_uri not on allow-listduring OAuth, add the exactredirect_urifrom the failing authorize request to the Upload-Post backend OAuth redirect allow-list. For ChatGPT clients this is typically onhttps://chatgpt.com/.../oauth/callbackorhttps://chat.openai.com/.../oauth/callback. - OAuth redirect callbacks are pre-allowed for: Claude (claude.ai/claude.com), ChatGPT, Cursor, VS Code (stable + Insiders), Smithery, Glama, Toolhouse, Perplexity (standard + Enterprise), Mistral Studio debugger and Postman — plus any
http://localhost/loopback redirect (RFC 8252), which covers Claude Code, Windsurf, Cline, Continue, Goose, Gemini CLI and othermcp-remote-style clients. Platforms without a documented fixed callback (e.g. Grok, Le Chat production) are added on request. - All tools declare MCP
readOnlyHint/destructiveHintannotations so clients can surface confirmation prompts for destructive operations, and the tools that publish or message people also require an explicitconfirmed: true(see Confirmation before acting).
Report a security issue: [email protected] (encrypted PGP available on request).
License
MIT © Upload-Post
