The slow part of self-hosting is not Docker. Every tool on the list makes you register your own developer app per platform, and Meta, YouTube and TikTok reviews can take weeks. Start those first.
Self-hosting a social media scheduler? Ask first what has to be awake when a post is due. Postiz, Mixpost and TryPost need their server up. CogSend runs on Cloudflare. pendpost splits by platform. https://t.co/KGarIclixh
Hardware differs a lot. Postiz's own docs: 2 GB RAM floor, 8 GB recommended for a team, plus Postgres, Redis and Temporal. pendpost: one Node process and a folder of JSON files. A backup is a copy of one directory.
A detail in pendpost I like: Fix with agent holds the post while the agent works. It may change that post's content, and its time only if it is overdue. It can never add a platform or touch the approval. The fence is server-side, not a prompt.
Renders eat disk. pendpost can delete local videos once they have posted. Off by default.
A video goes only once every post using it is done and a grace period has passed. Your agent can dry run it first.
The whole engine is MIT. The cloud only sells uptime.
You approved a post. Then its cover changed.
In pendpost that puts it back in Awaiting approval. So does an agent fix, or an edit to a post you approved weeks ago. Dragging it to a new time does not clear the hold.
Tokens in your own .env, dashboard on localhost.
Meta labels the error retriable:false. For a broken file that's true. For the timeout, the same bytes can pass on a lucky run. Diagnose first, then pick: fix the file, or switch to video_url.
Instagram API reel failing with ProcessingFailedError? We timed it: every rupload that passed answered within 26s, every failure came back as HTTP 400 at 35 to 37s, at any file size. Long reels run out of time. https://t.co/YaK4apMATJ
The other cause is the file. A reel whose audio track ended 5 seconds before its video failed every time. Compare per-stream durations with ffprobe, not the container duration, and re-encode if they drift more than a second.
What changed in the update? pendpost now tells you in the Studio: a short note per release, in English or Swiss German, once per install.
The notes ship in the package, so it works offline too.
An MCP server your agent can drive. You hold the approve button.
Every reply to you, in one list.
pendpost puts comments and replies under your posts, mentions, quotes, zaps, Google reviews and answers to your Radar replies in one place, newest first. You answer from the row.
Your agent reads the same list over MCP.
pendpost 2.12.0 is out. Approve a batch and the posts that can't go out stay behind, with the reason in one line. A blocked post shows its cause in a short line, fix beside it. Facebook turned off? Fix it on the card.
https://t.co/GzvIPUY8yB
Honest limit in the post too: an actor name is a label, not an identity. The gate stops an agent approving its own draft under its own name. Keep your client's prompt on for approve calls. MIT, all of it. npx pendpost
A human approval gate for an AI agent only works if it fails closed: a missing, stale or self-granted approval stops the publish. Six rules, a comparison table, and the real refusals pendpost returns. https://t.co/9fFixtZKwB
OWASP's LLM06 advice is to do authorization in downstream systems rather than relying on an LLM. A chat client's confirm prompt is a good first layer. The decisive check belongs where the publish happens.
Four replies waiting on one Bluesky login used to be four rows. Now they are one line in the Needs you tray: Bluesky not connected, 4 replies, Connect. Connect once, all four go out.
One count in the header, one filter in the list.
One command, no signup: npx pendpost