Limitations

What nicheclock cannot do yet, so you can decide before you invest time.

What it is not

Not a keyword tool. It reports that something is trending and why. It does not give search volume, keyword difficulty or rank tracking. Google Trends contributes an approximate traffic bucket for a trending search, nothing more.
Not a writer. It tells you what to cover and by when. It does not draft articles. A model, when configured, scores and explains; it does not generate content for your site.
Not a promise of traffic. A trending topic with a publish-by date is a lead, not a guarantee.

English and the US first. Keyword extraction, the bundled background corpus, the interface and the reports are English. Google Trends defaults to US, the Federal Register is American, and the Apple charts default to the US store. Per-country reading works in a basic form (Regions): only Google Trends and Apple are exact per country, Wikipedia and GDELT are approximate, and Hacker News, GitHub and your own feeds have no country at all. Country-specific keywords and region-aware Search Console gaps are planned (roadmap).
A fixed set of sources. Only sources whose terms allow it (Sources). Reddit, Product Hunt and similar are not read. You can add your own feeds.
Small niches stay quiet. A niche with narrow keywords may produce nothing for days. That is the relevance floor working.
Upcoming events come from limited places: Federal Register dates, feeds and calendars you supply, and Wikipedia early risers. ICS support covers simple recurrence rules only.
Index lag needs data. publish_by uses default_lag_days until Search Console gives at least three recently published pages to measure.

schedule: is not read. The field is stored in the config. The Docker image runs on an interval, and outside Docker you use cron or a systemd timer.
No built-in login on the report server. nicheclock serve serves static files with Python's HTTP server. Use a reverse proxy for authentication and TLS.
Not on PyPI, no prebuilt image. Install from a source checkout and build the image yourself until launch.
Tested on Linux. The test suite and CI run on Linux. macOS and Windows may work and are untested.
The report is plain. A page of niche tabs and tables with no search or filtering. feed.json is the interface for anything richer.
SQLite, single process. One store per config. Running two run processes against the same data directory at once is not supported.

One Search Console property per project.
The OAuth consent has to be done on a machine with a browser; it cannot be completed inside the container.
With the consent screen in Testing status Google expires the token weekly (details).

Gemini and Codex providers are unit-tested but not yet run against the real services.
A small local model judges less reliably than a large hosted one. Compare with nicheclock eval export before you rely on the ranking.
Model output is not reproducible: the same input can be judged a little differently on another day. Judgements are cached for a week to keep rankings steady.
Cost figures are estimates from a price table with a date; your provider's invoice is authoritative.
Items from publishers whose terms forbid AI use are never sent, so they keep their rule-based score only.

Free and open to anyone with a GitHub sign-in key, limited per key to 1,000 requests a day and 60 a minute, with no uptime promise. See API.
It serves only data licensed for redistribution, so it has fewer sources than a self-hosted install.
No Search Console data, no bring-your-own-model.
Custom niches from a URL are built by a daily batch, not live.