The agent watches your sources, filters out repeats and noise by itself, and sends you only what matters — and publishes what you approve
This is a one-agent newsroom on your own server. It walks your sources — websites, feeds, Telegram channels, search, product pages — collects material, merges repeats of the same story into one, and processes it with AI: a summary, an importance score, topics, companies and people, stories, translation. It then raises alerts by your rules, sends morning digests, and prepares publications that you approve. The important difference: the interface has no editing forms at all — you observe and give commands to your AI agent (a chat with Xedant Agent), and it changes the settings; if the chat is not connected, you can copy a ready command and send it directly, without the interface. Everything lives on your servers, and the settings are plain text files under version control. And when something is missing, any step can be extended with your own Python code: your own news collector, your own AI provider, your own delivery channel or publishing target.
The agent runs the news itself: you watch and give commands while your AI agent changes the settings
Fifteen ways to follow sources: websites, feeds, Telegram channels, search, product cards, podcasts, video and your own Python collector
One careful collector: no more than one request per minute to a site, and a repeated request only if the page changed
Repeats of the same news merge into one card that keeps every source address
Raw pages are saved in full, so you can see where every fact came from
Monitors watch specific values (price, availability, title) rather than the whole page
Every value has a history and a chart, and every change has before-and-after snapshots
An event appears only with proof: what it was, what it became and by how much
A summary strictly from the material, with links to specific quotes
An importance score from 0 to 10 and clear bands: noise, worth knowing, material, existential
Topics, companies, people and products are found automatically, and material about one event gathers into a story
Rules without programming: a tree of conditions by sources, tags, words and importance
Noise protection is on from the start: no more than three alerts per hour, quiet hours and a pause on repeats
Every suppressed alert is recorded with its reason, so the silence is always explainable
One morning digest instead of a stream of messages, on a schedule and by your topics
Eight delivery channels: email, Telegram, Slack, Discord, Microsoft Teams, a webhook, an API call and your own code
A channel sends nothing until you check it with a test message
A publishing draft is prepared from the material but goes out only after your approval
Invented quotes are dropped, and a draft without confirmation is not created at all
Ready feeds for your readers in three formats, protected by a token and only for published material
Everything on your own server, and the settings are ordinary text files under version control
Your own Python code extends any step: collection, processing, alerting, publishing — without touching the product itself
Sources: fifteen ways to follow the news
Feeds, ordinary pages and lists, search by query, a site map, public Telegram channels, ingest by email, document files, Reddit, YouTube, podcasts, Fediverse, reviews and your own Python collector — pick the ones where your news actually lives
You do not have to hunt for the feed by hand: the agent finds it from the site address itself, including the familiar conversions for YouTube, GitHub, Reddit and Mastodon
The list of sources moves in one file from any feed reader program (a list file those programs understand) and exports back the same way
Every source has its own schedule, its own headers and even its own connection channel, and its state is shown in words: Active, Failing streak, Error, Paused
Polite collection and a raw-page archive
All sources go online through one careful collector: by default no more than one request per minute to a single site, so no one blocks you for overdoing it
A repeated request goes out only if the page changed: when the site answers that nothing changed, the page is not downloaded again
Every downloaded page is saved in full, and identical texts are not duplicated, so an event always shows where the fact came from
A deleted page can be restored from an external web archive, and a short feed teaser is extended to the full article text
Repeats: one story — one card
The same news from different sources does not turn into ten records: repeats merge into one card
Merging goes by three signs in turn: the external item number, the page address and matching text within 72 hours
A merged record keeps every address it came from, so you see who wrote first and who picked it up
A short feed teaser is extended to the full article text when the page is open; when it is closed, what was already there stays, with no error
Monitors: what exactly changed
A monitor watches chosen values rather than the whole page: price, availability, title and any other field
Fields are described in words and the agent shows what it found, so you do not have to take the page apart by hand
Every value has a history and a chart, and every change has before-and-after snapshots with a storage period of its own
A trigger threshold keeps you from being pinged over pennies, and the baseline can be reset at any moment
Events with evidence
An event appears only together with proof: what it was, what it became, the absolute and the relative difference
Every event carries quotes and a link to the archived copy of the page, so you can check it yourself
The event card shows the change line by line and every delivery sent for it
An event can be marked as read or hidden from the feed; hiding affects visibility only and deletes no data
AI processing: summaries, importance, topics
A summary is built strictly from the material and backed by links to specific quotes
When the text gives no support, an honest note is put instead, and the AI invents nothing
Importance is scored from 0 to 10 and laid out in bands: noise, worth knowing, material, existential
Topics are found automatically, and an unprocessed event is honestly marked as Not scored instead of getting an invented number
Stories and entities: what connects to what
Material about the same event gathers into a story with the states New, Developing, Fading and Archived
The number of independent sources is counted, so you see whether one site wrote the news or dozens picked it up
Companies, people and products are recognized automatically and gathered into a register of entities with a mention history
Translation of material into other languages lets you read foreign sources without third-party tools
Rules: your own event filter
A rule is built from conditions without programming: by monitor, source, tags, topic, importance, words and entities
The check order is fixed and every decision goes into a log, so you can see in words why a rule did not fire
A rule can be tested on a past event before switching on, so nothing breaks in live work
A rule can mark, hide, send to a digest, start a publication and call a webhook
Alerts and noise protection
Every importance band has its own channels: existential goes to Telegram while worth knowing goes into the digest
Noise protection is on from the start: no more than three alerts per hour and a 30-minute pause on a repeat from the same monitor
Identical changes come once every 24 hours, and quiet hours from 23:00 to 08:00 put messages off until morning instead of throwing them away
An alert can wait for confirmation by a second source, and every suppression is recorded with its reason
Digests instead of a stream of messages
Instead of ten messages in a row there comes one digest, for example a morning digest on a schedule
For every rule you choose send now or into the digest, so something important does not drown in the common selection
Digests are gathered by topics and sources, and the schedule is tuned to your working day
A manual run does not shift the schedule, and a test send goes through the live channel, checking what you actually use
Delivery channels and the send log
Eight kinds of channels: email, Telegram, Slack, Discord, Microsoft Teams, a webhook, an API call (a request straight to another program, with no interface) and a channel that sends the message with your own Python code
Until a channel is checked with a test message it sends nothing: protection against thinking it is set up when it is not
The delivery log keeps the status, the number of attempts, the server answer, the request and answer texts and the chain of retries
Sending to internal and service addresses is forbidden, and passwords and tokens in the settings are masked
Publishing with a human check
A draft is prepared from the material of an event or a story in your style: language, tone, length and a sample text are set
Every quote is checked on the server: invented ones are dropped, and a draft without confirmation is not created at all
Nothing is published without your approval, and every rework is a new version that keeps the history
Six publishing targets: a Telegram channel, email, an RSS feed, a webhook, a folder with files and a target that sends the publication with your own Python code; unpublishing is described honestly for each channel
Feeds for your readers
A feed is prepared for your readers in three formats: RSS 2.0, Atom and JSON Feed
Only published material gets into the feed, so a reader never sees drafts or rejected items
The feed address is protected by a token: a wrong or missing token answers not found instead of showing the content
The token can be replaced in one action if the old one went to the wrong place
Observation screens and commands to the agent
Eight observation screens: Home, Monitors, Events, Inbox, Published, Delivery, Analytics and Scripts
Not a single editing form: changes are made by your AI agent while a person reads and gives commands
A command button puts a ready phrase into the chat with the agent, and without a chat it copies a ready command to send directly
New events, material and sends appear in real time, and states are labeled in words rather than codes
Analytics and archive search
You see how much news arrived, how much turned out important and how much noise the protection suppressed
AI spending is counted by task, and a daily quota keeps you inside the budget
Source quality is compared: who gives fresh news and who makes noise and repeats itself
Search runs across the whole archive and highlights what it found, and data sets can be exported for further work
Reliability: settings in files and control by API
All settings are text files in the project folder under version control: moving to another server is copying the folder
Changes are re-read on the fly, and a broken file does not break the work: the last good version stays with a clear error
The agent gets a separate management key, while the browser in the interface can only read and set marks
A self-documenting reference gives the commands, the descriptions of the settings files and eleven ready playbooks for the agent
Your own scripts: extend any step with your code
Your own Python code can be a source, an AI provider, a delivery channel or a publish target — or simply transform data at any step
A script is an ordinary text file next to the other settings: your agent writes and edits it, nothing to install by hand
Every run leaves a record: what was called, why, how it ended and how long it took — you can always see why a script stayed silent
Guardrails: a time limit, a cap on simultaneous runs and one global switch that stops every script at once
Every News Agent feature with no limits for individuals: unlimited installations on any of your computers and servers. $197 one-time payment for a lifetime license. 12 months of free updates, then (optionally) a 50% discount on the next 12 months of updates from the original purchase price. The matching Xedant Agent license is included — that is the chat with the AI agent you use to command the news.
Every News Agent feature with no limits for everyone in your company or team, external contractors included. $497 one-time payment for a lifetime license with 12 months of free updates, then (optionally) a 50% discount on the next 12 months of updates from the original purchase price. The Xedant Agent license for team work with the agent is included as well.
Unlimited News Agent instances under your own domain plus the full source code with the right to change it any way you like: restyle it, rename it for your company, integrate it tightly with your infrastructure. Anyone visiting your domain can use it as part of your services. $970 one-time payment for a lifetime license with 12 months of free updates, then (optionally) a 50% discount on the next 12 months of updates from the original purchase price. For smooth updates we recommend changing only the name and the integration — that way merging with new versions stays easy.
To evaluate the product calmly, a free trial license is issued for 30 days. In features it is no different from the paid one: the license restricts only sending messages to the chat with the AI agent, while everything else — collection, monitors, events, alerts, publishing, feeds, analytics and direct management — works without it. The trial key is issued on request — write to us on Telegram with your name and email.
Honest answer: the license restricts exactly one thing — sending messages to the chat with the AI agent. News collection, monitors, events, alerts, publishing, feeds, analytics and direct management all work without a license, so the product can be run by hand, without Xedant Agent.
What the license gives: the right to use the product, 12 months of free updates and support, and then a 50% discount on update renewals. The product is meant to work in a closed environment, so the license never stops it from working — it is about the right to use and to update, not about a key without which nothing starts.
