Post to social media from your AI agent
PlanVortex is a social media publishing platform with a public API and an official MCP server. An agent can schedule posts, read the comment inbox and answer direct messages across thirteen networks, and ask our planner to write the whole week. One command to install, and it authenticates with your own account's credentials.
npx -y planvortex-mcpThe official MCP server. Nothing to install: your AI client starts it.You need a PlanVortex app, and every plan has them: 1 on Free, 2 on Basic, 5 on Pro and up to 10 on Custom. You do not have to pay to connect an agent.
Published in the official MCP registry, on Glama and on Smithery.
What an agent can do with PlanVortex
Twenty-eight tools, grouped by what they act on. Nineteen read and nine write. The twenty-ninth, create_ai_plan, is not in the list until you switch it on, because it spends AI credits from your account.
| Group | Tools |
|---|---|
| Context | list_organizationslist_accountsget_plan_useget_unread_counts |
| Publishing | list_publicationsget_publicationcreate_publicationupdate_publicationretry_publication |
| AI planner | get_planner_templateslist_ai_plansget_ai_plan(create_ai_plan) |
| Media | upload_media |
| Comments | list_commentsget_comment_threadreply_to_commenthide_commentmark_comment_read |
| Messages | list_conversationslist_messagessend_message |
| Numbers | get_dashboard_summaryget_publication_statsget_top_publicationsget_account_metrics |
| Catalog | get_social_limitsget_social_capabilitiescreate_connect_link |
On top of that: three ready-made prompts —weekly_plan, inbox_triage and publish_from_brief— and four resources carrying each network's limits, its capabilities, the comment matrix and your organizations. The differences between networks travel as data, so the model never has to guess what each one allows.
The thirteen networks
Facebook, Instagram, Threads, LinkedIn, TikTok, X, WhatsApp, YouTube, Google Business, Bluesky, Discord, Telegram and Slack. Eleven of them publish; WhatsApp and Google Business have nowhere to publish to —one is messaging, the other a listing with reviews— so on those the agent reads, replies and measures. Slack is the other way round, and it is worth saying: it publishes and it measures, but it has no comments and no direct messages, because it is a team channel and not an audience.
The call that writes and publishes at once
Almost every social media MCP server lets the model write the copy itself. This one does not need to: PlanVortex has its own planner, and an agent can ask it for a full week built from a theme, from your own photos, from an article or from your shop's catalogue.
> Plan the week for the bakery: sourdough bread, use my August photos and post it to Instagram and LinkedIn. get_planner_templates() → from_images: 0 image credits · 7 posts create_ai_plan({ template: "from_images", accounts: [...] }) → budget: 350 credits · status: pending get_ai_plan(plan) // minutes later → generated · 7 drafts with copy and photo
Generating takes minutes and is not a single call: create_ai_plan queues the plan and returns the budget, get_ai_plan is polled until it finishes, and what comes out are drafts. Scheduling them is still your call, one post at a time.
The cost is known before anything is generated
The server works out what the plan costs and returns it when queuing. A plan that does not fit your credits is rejected whole instead of half-generated, so an agent stuck in a loop cannot drain your account without you seeing it.
And you switch it on, with one line
create_ai_plan is the only tool that spends money, so it is registered only with PLANVORTEX_MCP_ALLOW_AI=1 in your configuration. Switched off it does not even appear in the tool list: there is no way to call it. A person writes that line once, before any agent starts.
Five templates, and one spends no images
get_planner_templates publishes all five with what each one costs, so the agent can explain the price before asking your permission. The one that starts from your own photos spends no image credits at all.
How to install the MCP server
Three steps, and none of them goes through us: the server runs on your machine, with your credentials.
Create an app in the panel
Under Settings → Apps. It gives you a client_id and a client_secret. Apps are on all four plans, so this costs nothing on the free one.
Add the server to your AI client
One command in Claude Code, or the JSON block below in Claude Desktop, Cursor and VS Code. Nothing to install: npx downloads it on start.
Ask it for something
"What do I have scheduled this week, and which comments are still unread?" If it answers with your real accounts, it works.
claude mcp add planvortex \ --env PLANVORTEX_CLIENT_ID=... \ --env PLANVORTEX_CLIENT_SECRET=... \ -- npx -y planvortex-mcp
{ "mcpServers": { "planvortex": { "command": "npx", "args": ["-y", "planvortex-mcp"], "env": { "PLANVORTEX_CLIENT_ID": "...", "PLANVORTEX_CLIENT_SECRET": "...", "PLANVORTEX_ORGANIZATION_ID": "..." } } } }
PLANVORTEX_ORGANIZATION_ID is optional and worth setting: without it, the agent spends one call per conversation working out which organization you mean.
No Node installed? Every release also ships a .mcpb bundle that Claude Desktop installs with a double click.
Download the latest .mcpb →Every environment variable
| Variable | Required? | What it does |
|---|---|---|
| PLANVORTEX_CLIENT_ID | yes | Your app's client_id. Every plan has apps. |
| PLANVORTEX_CLIENT_SECRET | yes | Its secret. Never passed as a tool argument. |
| PLANVORTEX_ORGANIZATION_ID | no | Default organization. Saves a discovery call per conversation. |
| PLANVORTEX_BASE_URL | no | Point at another PlanVortex deployment. |
| PLANVORTEX_MCP_UPLOAD_DIRS | no | Directories upload_media may read from. Empty means none. |
| PLANVORTEX_MCP_AUTH_TOKEN | with --http off-loopback | Bearer token the HTTP endpoint requires to bind anywhere but loopback. |
| PLANVORTEX_MCP_READ_ONLY | no | 1 removes the nine write tools from the listing. |
| PLANVORTEX_MCP_ALLOW_AI | no | 1 adds create_ai_plan, which spends AI credits. |
| PLANVORTEX_MCP_LOG_LEVEL | no | debug, info, warn, error or silent. Always to stderr. |
What it cannot do, and how to switch off what it can
This server puts text written by strangers —a comment, a review, a direct message— into the model's context while that same model holds tools that publish to your accounts. The limits below are the answer to that, and the first one is not a setting.
It deletes nothing, and that cannot be turned on
No tool deletes a post, an account or a comment. It is not an option switched off by default: the code does not exist. The worst possible case has to be something you can see and undo.
Read-only, for real
With PLANVORTEX_MCP_READ_ONLY=1 the nine write tools are not registered, so they are not in the listing and the model cannot even try to call them. Nineteen read-only tools remain.
It runs on your machine
The client_secret lives in the process your AI client starts, never on a server of ours. And if you would rather host it yourself, --http mode ships its own Dockerfile, listens on loopback only and demands a token to bind anywhere else.
It does not read your disk
upload_media only opens files inside the directories you declare in PLANVORTEX_MCP_UPLOAD_DIRS, and that list is empty out of the box. Without configuring it, the agent cannot upload a file of yours even if the model asks.
If your agent does not speak MCP, it speaks HTTP
MCP is the comfortable door, not the only one. Everything the server does is also in the public API, which is open on all four plans and demands no paid plan to integrate.
The API, 135 operations
OAuth 2.0 with client_credentials, JSON, and the same numbered error catalogue the libraries use. The full documentation is published and public.
Read the documentation →Two official libraries
npm i planvortex and pip install planvortex. Both publish with provenance and carry the token manager, file uploads and typed errors.
For developers →And four files you can read without crawling anything
They sit at the site root, they are plain text and there is no JavaScript in front of them. If you are an agent evaluating this tool, start here.
/llms.txtWhat PlanVortex is, who it is for, with verified prices/llms-full.txtThe whole API in plain text: all 135 operations with their parameters/openapi.jsonThe complete OpenAPI, to generate a client in any language/pricing.mdThe tariff, in markdown, with no form to fill in
Frequently asked questions
What is the PlanVortex MCP server?
It is the official Model Context Protocol server connecting an AI assistant to your social media accounts. It gives twenty-eight tools to schedule posts, read comments, answer messages and pull metrics across thirteen networks. It installs with npx, runs on your machine and uses the credentials of an app in your PlanVortex account.
Do I need a paid plan to connect an agent?
No. The apps that give you the credentials are on all four plans, the free one included, which comes with 1 app, 3 connected accounts, unlimited posts and 1 GB of storage, no card required. What changes with the plan is how many separate apps you get and the call rate, not whether you can integrate.
Which social networks can an agent publish to?
Eleven: Facebook, Instagram, Threads, LinkedIn, TikTok, X, YouTube, Bluesky, Discord, Telegram and Slack. WhatsApp and Google Business complete the thirteen, but you do not publish on those: they are messaging and a review listing, so there the agent reads, replies and measures.
Can an agent delete a post or an account?
No. None of the twenty-nine tools deletes a post, an account or a comment, and it is not an option left switched off: that code does not exist in the server. It is the only serious guarantee you can give a model that reads strangers' text while holding the ability to publish in your name.
Where does my client_secret end up?
In the process your AI client starts, on your machine. It never goes through a server of ours and it is never passed as a tool argument, so it does not end up in the conversation history or in the model's context.
Can the agent write the content, not just schedule it?
Yes, and the model does not write it: the PlanVortex planner does, from a theme, your own photos, an article or your shop's catalogue. Generating spends AI credits, so that tool has to be switched on by hand with PLANVORTEX_MCP_ALLOW_AI=1. What comes out are drafts.
Does it work with Claude Desktop, Cursor and VS Code?
Yes. Claude Code is configured with one command and the other three with the same JSON block in their MCP servers file. If you do not have Node, every release also ships a .mcpb bundle that Claude Desktop installs with a double click.
What if my agent does not use MCP?
Then it uses the public HTTP API: 135 operations over OAuth 2.0, open on all four plans. There are official Node and Python libraries, the complete OpenAPI published at the site root, and a plain-text version of the documentation meant to be read by a model.
One command, and your assistant publishes
Create an app on the free plan, paste the two credentials into your AI client and ask it what you have scheduled this week. No card required.