Job postings within hours of going up, fetched and filtered on your own machine
Run it when you like. It reads the boards you choose, and a language model reads a posting once for the facts your rules check. What you can't take is dropped with its reason, and the rest reaches you hours old, not days later with a hundred applicants ahead of you.
Check your inbox.
A confirmation link is on its way to . Open it and press Confirm to join the list. The link works for 7 days. It can take a minute to arrive, and it may land in spam.
The local app, sheet by sheet
The dashboard, Sent view. Eight of the newest sent postings from public boards, picked: the title, the board, when it was posted and when a run saw it. Real rows from my run, company names left out. The chips count the whole run so far under the 30-day range; the page's week is an average over two weeks of it.
The dashboard, Dropped view. What was dropped, with the check's reason code as the status chip. Eight of the nine rows shown below, picked one per reason, as the dashboard draws them.
The dashboard, Geo view. The held pile: postings the model called remote that hire only in one or two named countries that leave mine out, each with its places on a geo chip, kept for a look and never sent. Seven real rows, picked.
Filtering. Freshness (the age limit), Location (work types), Hiring scope (how many regions a limited remote posting must name), the pay floor the Model step checks, and whether a likely ghost counts as dropped.
Search. Keywords: the stack words a page must hold and how many of them, the search terms for the sweep of company career pages, and the stop words that drop a posting by its title, matched as whole words, so Java never catches JavaScript.
Blocks. Blocklist (blacklisted companies), Language (the languages you read) and Location (the places you can and can't work from; the allowed places are also your country for the Hiring scope check). The values are an example, a search from Poland.
Model. Your OpenRouter key and the model, from a tested list, with a button that checks the connection. With the key left blank, postings wait in the queue.
Prompt. The one prompt the Model step sends, asking every question about a posting at once, and the shape of the answer. It's a text box, so the questions are yours to change, and the {{areas}} slots are filled from your own list of role areas.
Notifications. Where the Digest goes: a Telegram bot, an email account through SMTP, or both.
Job boards. What Read walks: one listing page per line, the first lines of my list of about thirty here. A line starting with // is switched off, and Ctrl+/ toggles the selected lines, as in an editor.
What it dropped, and why
In a week, PostingWatch checks 4,915 new postings for me and sends me 292. Here are nine of the others, one per reason.
From the dashboardNine of the 4,915 postings in a week
- Mid-Senior Frontend Engineer (Saudi National)ModelRequires citizenship
- Senior AI EngineerLocationJudged not remote by the model
- Developer - Fullstack (REACT/Java/Python)LocationOn my excluded-locations list
- Senior Software Engineer (Go - React - TypeScript - Next.js)LanguageWritten in Spanish
- Senior Technical Lead - Java/ReactKeywords“Java” in the title, a word I excluded
- DevOps EngineerKeywordsNo keyword from my stack on the page
- Senior Software Engineer (Frontend), Business AccountsFreshnessPosted 6 May, seen 1 October: past my 3-day limit
- Software Developer, React Native - HybridFingerprintThe same text as a posting already on record
- Software Engineer, ProductHiring scopeRemote, but hiring in the United States only. Kept in the dashboard, not sent.
A week is the average of two weeks of my runs, ending 2 October 2026; every reason with its exact count is in /llms.txt.
How a posting gets checked
A posting stops at the first check it fails, and that check is its reason in the dashboard, next to its saved text. The ghost check only flags.
Read
DroppedA run opens two browser windows, walks every board in your list, reads each new posting and closes them when it's done. A blocked or empty page is dropped with that reason and kept in the dashboard, so you can open it yourself; a bot wall raises a notification, and each board's health is on the dashboard.
Rules
DroppedBefore any model call, grouped by check, not in the order they run:
- Keywords: a word from your stack on the page, and no stop word in the title, matched as whole words, so Java never catches JavaScript.
- Location: the board's work-type badge matches yours, the place isn't one you excluded, and a hybrid or on-site role is somewhere you allow. The model's answer can drop a posting here too: judged not remote, or refusing applicants from outside its countries.
- Freshness: not closed, under 100 applicants, not marked ineligible by the board, and within your age limit.
- Blocklist: not a company you blocked.
- Fingerprint: not the text of a posting already on record.
- Language: one you read.
Model
DroppedA language model reads the posting's title and text once and reports facts, all in one answer:
- is it a job page
- is it really remote, and for where
- what area and main stack it is for
- does it ask for a language you don't speak
- does it require citizenship or a clearance
- is the stated pay under your floor
The area and stack rule is tuned to my own search, frontend and JavaScript full-stack, in the code rather than the settings for now. No answer means a retry, not a drop.
Hiring scope
HeldThen, from the same answer, a remote posting naming one or two countries or regions that leave yours out (your allowed locations, in the settings) is held: out of the digest, shown apart in the dashboard with a geo badge, worth a look when a company might make an exception. Three or more named regions pass, as the table below shows.
Ghost check
FlaggedEvery posting gets this check, whatever the others said. A posting closed within days of going up, listed again by the same company on the same board, marked reposted or always open, or from an agency or a board hiring for someone else gets a likely-ghost flag in the dashboard and stays where it was.
Digest
SentWhat passed goes out in one email or Telegram message, grouped by board, once the run's queue is done. A board that blocked the reader is named in it.
Remote postings, case by case
What the Location and Hiring scope checks do with a posting, for the example settings (a reader in Poland).
| The posting | Check | Outcome |
|---|---|---|
| On-site or hybrid, anywhere | Location | Dropped |
| In a place on my excluded list, such as Bangalore, India | Location | Dropped |
| Called remote, but the model finds in-person time after hire | Location | Dropped |
| Remote, but refusing applicants from outside its countries | Location | Dropped |
| Remote, worldwide, including through an employer of record | Hiring scope | Sent |
| Remote, no place named | Hiring scope | Sent |
| Remote, naming my country or a region it is in: Poland, the EU, EMEA | Hiring scope | Sent |
| Remote, naming three or more regions that leave mine out: US, UK, Canada | Hiring scope | Sent |
| Remote, naming one or two countries that leave mine out: US only | Hiring scope | Held |
What arrives in your inbox
One email or Telegram message when new postings pass, grouped by the board they came from. Each line has the title, the company, the location and the day it was posted; the exact time, when the board gives one, is in the dashboard.
PostingWatch · 4 new jobs
Greenhouse boards
- Senior Frontend Engineer, React
- Staff Engineer, Web Platform
Ashby boards
- Frontend Engineer, Design Systems
Hacker News: Who is hiring
- Founding Frontend Engineer
It runs on your machine
- Your keys
- Model calls go to OpenRouter with your own key, or to a local model through Ollama. The model sees a posting's title and text with your languages and pay floor, and nothing else about you.
- Your data
- Postings, reasons and settings stay in a SQLite file and a settings file on your disk. There's no account, no telemetry and no server of mine in between.
- Your runs
- A run is one pass over every board. You start each one yourself, from the dashboard or a terminal, while your machine is awake; there is no schedule. The digest goes out when the run's queue is done.
Questions
What do I need to run it?
At release: download it, run one command, and paste your OpenRouter key. It needs Node.js and a Chromium-based browser, which it opens itself to read the boards, and an email account (a Gmail app password works) or a Telegram bot for the digest. A local model through Ollama works in place of OpenRouter. Runs start by hand; there is no schedule. Windows and macOS support is in progress.
Could it drop a job I'd have wanted?
It can. The posting is then in the dashboard under the check that dropped it, with its saved text, so you can see what the rule or the model made of it and loosen the rule. No answer from the model keeps a posting in the queue for a retry rather than dropping it. My own rules are wide on purpose, so that it errs toward sending: about 3 in 10 of what it sends me are worth applying to, and I take that over a lost posting.
Which boards does it read?
The ones in your settings, one listing page per line. My list has about 30: company career pages on Greenhouse, Lever, Ashby, Workable, SmartRecruiters and Workday, Hacker News, We Work Remotely, Remote OK, Wellfound, Working Nomads, Built In, Y Combinator's boards, arc.dev and more. Another listing page is one more line; a board with unusual markup needs a small reader of its own. A board that shows postings only to signed-in members is read in a browser profile of its own, which you sign into once.
Does it apply to jobs for me?
No. It reads new postings, drops the ones you can't or wouldn't take, and sends you the rest. Deciding what to apply to stays with you.
What does it cost to run?
The code will be open source, and the model calls cost what your provider charges. On my setup a run costs $0.01 to $0.06; the first few cost more, before anything is cached. My OpenRouter key spent about $0.45 a day over about a dozen runs a day, read from the key on 2 October 2026, so about $14 a month at my pace. A local model through Ollama costs nothing beyond your own machine.
When will the code be public?
There's no date, and I won't invent one. It is the tool I run every day for my own search, and the public repository comes once it installs cleanly on a fresh machine. Leave your email above and you'll get one message when it opens.
How is it related to PostingFit?
PostingFit is a separate paid product, still being built, that tailors a resume to one posting. PostingWatch will be open source either way, will be able to hand a posting to it, and works without it.
Get one email when the code is public
PostingWatch will be open source. The form at the top is the whole signup: one address, one confirmation, one email when the repository opens.
Go to the form








