Ziba

Ziba is a self-hosted personal content aggregator. It collects articles from configured sources, processes them with AI, and presents them as a personalized magazine.

Sources and interests are both configured by hand. An article is collected from its source and appears in the last-24-hours digest only if it matches an interest and scores above a configured threshold. Articles that pass are then summarized.

The supported source types are:

• RSS feeds
• Email newsletters

In the web interface, articles are grouped by interest and ordered by score.

Scoring and summarization use models chosen by the operator: a fast one to analyze and score each article, and a more capable one to write the summary. Models from OpenAI and Anthropic are supported.

Each article goes through two stages. Assessment says what the article is about and rates it against the interests, in a single call on the fast model. Summarization writes a summary aimed at those interests, on the capable model, and only for articles above the threshold.

Articles below the threshold keep their score and stay browsable in the archive. They are simply not summarized and not promoted: the AI curates, it does not censor.

The reader can mark any score as “Should be higher” or “Should be lower”. Ziba collects those corrections, but does not currently use them to adjust scores. Feedback is never sent to the AI provider, the provider's original score is preserved, and Settings → Scoring can remove all feedback and restore scores from their provider values.

An offline analyzer is also available. It only matches against the configured interest vocabulary, so it is crude — it exists for debugging, not for daily use.

Deploy

Use the compose.yaml file available here in the repository.

Compose pulls ghcr.io/emaori/ziba for the application and ghcr.io/emaori/ziba-browser for the Firefox feed helper. The third service is Postgres. The browser has no published port and is used only by RSS sources that explicitly opt in.

Fill in the variables of the compose according to the documentation in this table (and in the compose file iteself).

Setting	What it decides	What to put
POSTGRES_PASSWORD	the database password	change it — anything
ZIBA_DATABASE_URL	how Ziba reaches the database	the same password, after ziba: — it ships blank
ZIBA_BROWSER_URL	how Ziba reaches the Firefox feed helper	keep the Compose value http://ziba-browser:3000
ZIBA_AI_PROVIDER	which company answers	openai or anthropic
OPENAI_API_KEY	pays for it	your key — or ANTHROPIC_API_KEY, uncommented, if you chose anthropic
ZIBA_FAST_MODEL	scores every article	required, e.g. gpt-5.6-luna
ZIBA_CAPABLE_MODEL	writes the summaries	required, e.g. gpt-5.6-terra
ZIBA_FAST_EFFORT	how hard the fast model thinks	low — raising it costs more than the article does
ZIBA_CAPABLE_EFFORT	the same, for summaries	low
TZ	which local clock the schedule follows	your zone, e.g. Europe/Rome; otherwise UTC

On the first launch, open Ziba in a browser. The setup wizard has three steps: interests, sources, then the collection schedule. Add or edit one item on its own page. You can also add a preconfigured interest or RSS source, then edit it if needed. The Schedule step proposes the defaults 6h and 04:00:

• Run every is the interval between runs
• Start the daily cycle at anchors those runs to the local time set by TZ. For example, 6h starting at 04:00 runs at 04:00, 10:00, 16:00, and 22:00. Enter 0 as the interval to stop scheduled collection. Changes take effect without restarting Ziba.

It also proposes starting the first collection as soon as setup finishes; clear that checkbox to wait for the next scheduled time. Collection does not start until setup is complete.

Newsletter credentials can be entered or changed in Settings. Stored usernames and passwords are never displayed again.

Turning off collection for a source stops future collection. It does not remove articles that Ziba already collected.

An RSS source that receives 403 responses from an ordinary HTTP client can enable Fetch this feed with Firefox. Use it only for affected feeds: it is slower and requires the browser service. Firefox downloads the feed XML only; article and roundup pages continue through Ziba's ordinary HTTP client.

Things worth knowing

There is no login. Ziba is single-user by design and has no accounts, no password and no session. Whoever can reach the port is the reader, and can mark articles read as well as read them.

It costs money to run. Every collected article is assessed by the fast model, and everything above the threshold is summarized by the capable one. On one real archive of 485 articles that came to about $1.50 with OpenAI, or roughly $4–5 a month at fifty articles a day. Two settings move that figure more than anything else: the threshold in Settings, which decides how much reaches the expensive model, and ZIBA_FAST_EFFORT / ZIBA_CAPABLE_EFFORT, because reasoning tokens are billed as output and output costs several times input. The Statistics page reports exactly what has been spent, per interest and per day.

The same page shows whether retrieval or analysis is falling behind. Temporary processing failures retry automatically; repeatedly failing items are reported there instead of blocking newer work.

It runs without a key. With no API key configured, collection, retrieval and the archive all work, and only the analysis is skipped. That mode is crude and meant for debugging.

Newsletter credentials are write-only in Settings. Ziba never displays a stored username or password. Leave those fields blank while editing to keep the stored values. With Gmail, use an App Password and paste it without spaces.

Web collection only connects to public addresses. RSS feeds, roundup pages, article pages, browser-fetched feeds and every redirect are refused if their hostname resolves to a loopback, private, link-local or otherwise non-public address. This prevents an external feed or newsletter from using Ziba to reach services on the machine or home network. IMAP servers are explicitly configured by the operator and are not subject to this restriction.

Debugging what the model was asked

Set ZIBA_MODEL_JOURNAL=true and every request to a model, with its reply, is appended to log/modelJournal.txt, which the compose file bind-mounts so it can be read from the host.

Headers are never recorded, so the API key never reaches the file.

The container runs as nobody, uid 65534, so the log/ directory on the host has to belong to it. Docker creates that directory as root on the first up, and root's is not writable by nobody:

chown 65534:65534 log

Without it, startup fails with a message. It only matters while the journal is switched on — nothing writes there otherwise.