Social Daily Poster

Nobody has the time, or the will, to sit down and write a post every day. Social Daily Poster does it for you, from what you actually did that day.

Social Daily Poster is the open-source alternative to Taplio: it writes your LinkedIn posts from the work you already did, on your own machine, for the price of the AI subscription you already pay for.

What it does

Every night, Social Daily Poster looks back at your day: your coding-agent sessions (Claude Code or Codex/ChatGPT), screenshots, and optionally your Telegram, Gmail, WhatsApp, and even a spoken debrief you record after a call. It picks the single best story and drafts a post. That draft lands in a private Telegram bot with Approve / Edit / Skip buttons. Approve also draws an illustration and shows it to you before anything is public. Nothing posts until a second tap.

Once that post is live on LinkedIn, Social Daily Poster can spin off the same story for X (a real rewrite, not a truncation) and Reddit (a prefilled link you submit yourself), each with its own approval step and its own on/off switch.

Runs on one laptop for one person, no server needed. That's how most people use it. It can also run on a shared server for a few people.

Why it's good

• Posts come from what you actually did. They read like your week, written in your voice.
• Never auto-publishes, on any platform. Every channel needs an explicit tap after you've seen the exact text (and, for LinkedIn, the exact image).
• Runs on your own machine. In laptop mode nothing leaves it except what you explicitly approve for posting.
• Costs close to $0. Drafting rides on the Claude Code or ChatGPT subscription you already pay for (see Cost).
• Posts through real, official APIs: LinkedIn and X, not a workaround that risks your account. Reddit has no workable automated-posting API for a personal project right now, and this README says so plainly (see What it posts to).

Quickstart

Fastest: hand it to a coding agent. If you already use a terminal-based agent (Claude Code, Codex, etc.), open it in an empty folder and give it this:

Clone https://github.com/dimamak/sdp and set it up locally for one person.
Before running the wizard, ask me which channels I want: Telegram and
LinkedIn are required, then check whether I also want Gmail, WhatsApp,
images, X, or Reddit, don't just leave them off by default. If I don't have
a LinkedIn Page, use https://www.linkedin.com/company/sdp-page/ as a
placeholder. Then walk me through the wizard with those answers.

The wizard is interactive by design: you'll still need to answer things yourself, like pasting a Telegram bot token or approving a LinkedIn login in your browser. But the agent can drive the whole install and relay each question to you as it comes up.

Or run it yourself (works on Windows, macOS, Linux):

git clone https://github.com/dimamak/sdp.git && cd sdp
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt        # Windows: .venv\Scripts\pip
.venv/bin/python -m setup.wizard                 # mode? -> laptop (the default)

The wizard walks through: which sessions to harvest, your Telegram bot, and (all optional) Gmail/WhatsApp/LinkedIn/images/X/Reddit. Every step is re-runnable on its own, e.g. python -m setup.wizard --source linkedin. When it's done:

python -m server.pipeline.run_nightly --dry-run   # one harvest + digest, no posting
python -m server.bot.main                         # leave this running

That one process polls Telegram for your taps and schedules the nightly draft itself: no cron, nothing else to install. python -m setup.wizard --doctor checks your setup any time.

Want a shared, always-on server for a few people instead of one laptop? See Server mode.

What it captures

Source	What it captures	Notes
Claude Code / Codex / ChatGPT sessions	Your coding-agent transcripts	Read locally; covers Claude Code, the Codex CLI, and the ChatGPT desktop app
Screenshots	Whatever you drop into a watched folder	Any OS
Telegram, Gmail, WhatsApp	Your own message/mail history	All optional, all off by default; WhatsApp capture is read-only
Call / office audio	Local speech-to-text on recordings you make, in whatever languages you actually speak	Off by default; recording and transcription both run on Windows, macOS and Linux
Screen activity	Which app and window was in the foreground, plus occasional screenshots — the non-coding half of the day	Off by default; not supported on Wayland (see Limitations)

Every source is opt-in and walked through by the wizard. config.example.yaml has the full field-by-field detail as inline comments.

What it posts to

• LinkedIn: the official w_member_social API. This is the core flow: draft → approve → optional illustration → post. Creating the LinkedIn app needs an associated LinkedIn Page; if you don't have one, use linkedin.com/company/sdp-page, a placeholder page for this project that anyone setting up their own instance is welcome to use. It isn't tied to any real business and has nothing to do with what actually gets posted.
• X (optional, x.enabled): once a post is live on LinkedIn, the same day's story gets a genuine X-length rewrite (LinkedIn drafts run 1,100–1,600 characters; X's default cap is 280), with its own Approve/Rewrite/Skip step. Uses X's official API.
• Reddit (optional, reddit.enabled): Reddit closed off practical API access for a project like this. Social Daily Poster hands you a prefilled submit link and ready-to-paste title and body, and you review it and tap Submit yourself. Nothing is ever posted by code.

Skipping or failing a later step (X, Reddit) never touches a post already made on an earlier one. See Cost for what each channel actually costs to run.

X reply radar (optional, radar.enabled, off by default)

Discovery and draft-assist only — this never replies for you. A browser extension you install and scroll x.com with (free) surfaces posts worth replying to and, once you've told it enough about who to watch, an optional budget-capped API poll (default cap $5/month) does the same for a short watchlist. Either way, all you ever get is a drafted reply in Telegram — Approve · Edit · Skip — that you paste in and send yourself. No code path here calls X's post endpoint. See config.example.yaml's radar: block and python -m setup.wizard --source radar to turn it on.

Images

Approve doesn't publish right away. It first asks the same session that wrote the post for an image brief, renders it, and sends it back with Post with image · Regenerate · Post text-only · Cancel. Reply to steer it ("more abstract, no people") and each reply is a new take; regenerating never touches LinkedIn, only the final tap does. Turn it off entirely with image.enabled: false.

Every render is checked for broken text — a misspelled word, an invented brand, a legible dashboard or slide — before it's shown to you. A flagged take is silently re-rendered once (it never counts against your regeneration budget, since you never saw it); if the retry is still bad, it's delivered anyway with a warning in the caption rather than blocking the post. Configure under image.text_check*.

Publish window

By default, tapping Post with image / Post text-only sends it to LinkedIn immediately. Set publish.window (e.g. "15:00-18:00") and it queues instead, going out at the next eligible slot on its own — useful if you approve first thing in the morning but don't want every post landing at the same hour. Post now anyway on the queued message bypasses it for that one post. publish.days caps which weekdays are eligible; /status lists what's currently queued. Leave publish.window empty for the old immediate behaviour.

Privacy & safety

For a tool that reads your coding transcripts, this is the part that matters most.

• Everything stays on the machine you configure. In laptop mode, nothing ever leaves it except the exact text (and image) you tap to approve.
• The drafting/rewrite step runs the LLM in a restricted mode with no general file or tool access.
• WhatsApp capture is read-only by design.
• Audio is transcribed locally, and the recording is deleted once the transcript exists. The raw audio never persists.
• Recording other people's conversations has privacy implications of its own: don't quote anyone identifiably in a public post (the default style guide already forbids it).

See CONTRIBUTING.md if you want the exact sandboxing mechanics behind the Claude vs. Codex backends.

Cost

Item	Marginal cost
Drafting, image briefs, X/Reddit rewrites, session summaries	~$0; runs through your existing Claude Code or ChatGPT/Codex subscription's CLI (an API key also works if you'd rather pay per-token)
LinkedIn publishing	$0 (official API)
Post illustrations (optional)	~$0.13–0.24/render; image.enabled: false skips this entirely
X posting (optional)	Pay-per-use, no free tier; well under $1/month at one post a day
Reddit (optional)	$0; no API or account, just a link you submit yourself
X reply radar (optional)	Browser-extension lane is $0; the optional API polling lane is capped at radar.monthly_budget_usd (default $5/month, $0 disables it)

At one post a day with images on, expect low-single-digit dollars a month.

Server mode / several people on one server

For an always-on shared box instead of a laptop, with a separate laptop feeding it over SSH:

git clone https://github.com/dimamak/sdp.git && cd sdp
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python -m setup.wizard        # mode? -> server
.venv/bin/python -m setup.wizard --doctor

Each additional person gets their own instance: separate config, secrets, store, Telegram bot, and LinkedIn token, via .venv/bin/python -m setup.wizard --instance alice. People can share one LinkedIn or X App's credentials the same way a team shares any App: whoever sets it up first pastes the client id/secret, everyone else just runs their own login to get their own token (see the wizard's linkedin/x steps).

Full walkthrough (cron, systemd, Windows laptop push, CI auto-deploy) lives in docs/self-hosting/ci-deploy.md and the wizard's own prompts.

Limitations & non-goals

• Only local coding-agent history is read. Codex Cloud tasks and the plain ChatGPT web/desktop chat tab aren't in the local session store, so they aren't harvested.
• WhatsApp carries real, if low, ToS/ban risk: it goes through an unofficial client, not WhatsApp's Business API. Read-only by design.
• Reddit is draft-assist only, never automated posting (see What it posts to).
• The X reply radar is discovery-and-draft only, never automated replying (see X reply radar). Sending an LLM-drafted reply is still a suspension risk regardless of how it was drafted — the mitigation is editing drafts and keeping volume low, not this tool.
• Audio and screen capture need the bot process to be running. In laptop mode both recorders are threads inside it, so capture stops when it stops. The wizard registers an OS-native autostart (systemd --user, launchd, or a Task Scheduler logon task) so it comes back at login, and --doctor fails if the bot hasn't checked in.
• Screen activity capture doesn't work on Wayland. There's no compositor-independent way to read the foreground window, or to grab the screen without a portal prompt per shot. The recorder refuses to start there rather than logging ? all day; an X11 session works. On macOS, window titles additionally need Accessibility permission — without it you get app names only.

Contributing / Security / Licence

• Found a bug or want a feature? See CONTRIBUTING.md.
• Found a security issue? See SECURITY.md. Please don't open a public issue for anything that touches credentials or account safety.
• LICENSE: MIT.