How it works
The AI does the work. You say what you want.
Setting up is a conversation, asking is a sentence, and every answer shows how it was made. Here’s what happens at each step.
1 · Watch something new
One message, one approval.
Say “Watch Halden Mobile and their CEO.” Typestream researches them: Halden Mobile is the Swedish phone maker (not the Norwegian town, and not the prison), and its CEO is Dana Okafor. It sets them up as two subjects.
It checks the profile it wrote against a couple of hundred recent articles, reads the borderline ones itself, and tightens the profile until the judge agrees with it. It never asks you to sort articles.
Then it offers to fill in the last 30 days, with the price. Once you approve, each subject gets a dashboard (coverage volume, tone over time, the tier-1 share and the latest notable stories) and two quiet monitors: a first tier-1 story, and a sharp drop in tone.
- Halden Mobileorganisation412 articles this week−0.14
- Halden Aeroproduct190 articles this week−0.18
- Dana Okaforperson61 articles this week+0.05
2 · Ask
Any question, answered with a chart and its receipts.
Subjects. The chat works out who a question is about. Add one with @, or remove one.
Plan. It turns the question into what to judge for each article and what to compute. It checks what’s already answered and picks the cheapest plan that answers well: “Reusing tone scores for 3,904 articles, so this costs 12 credits instead of 190.”
Run. Anything under your auto-approve limit runs straight away; anything over it asks once, with the price. A sampled answer arrives within seconds, and the chart tightens as the full run completes.
Answer. A chart or breakdown, a short explanation with citations, and next steps: pin it, watch it, or show the articles. Follow-ups reuse the answers already in hand, so they cost less.
292 articles behind “Durability and fit, 71%”
- T1The LedgerEU regulators ask Halden about MagPort failures−0.42
- T2Repair JournalMagPort connectors are wearing out, owners say−0.61
- T3Pocket ReviewSix months with the Aero: what’s held up+0.08
Every answer has three layers
Nobody needs the third layer. Anyone can reach it.
- Layer 1
The answer
A sentence and a chart. Every number opens to the articles behind it.
- Layer 2
How was this made?
The method in plain words: the scope, each question as a sentence, the counts at each step, what was reused, and the credits spent.
- Layer 3
Details
The exact questions asked, the thresholds, the model version, and every article’s probabilities, with a CSV.
3 · Stay ahead
Live cards, and alerts written as sentences.
Say “pin this”, or click Pin. The answer becomes a card on the subject’s dashboard, and it recomputes as new articles are judged. Discuss opens a chat that knows the card; Edit changes it in conversation (“show it monthly”, “split by country”).
After a notable answer, the AI offers to watch it. A monitor is a plain sentence, the channels it tells, and its cost a month. Alerts go by email, Slack, Teams or webhook, each with a short summary, the articles that caused it, and Discuss.
Profiles keep themselves current. Once a week the AI rechecks fresh borderline articles, adds new names as they appear, and logs every change with its reason, so you can revert it.
Alert #aero-launch when tone toward the MagPort drops 0.2 below its 4-week average.
What the AI does
It does the work, and asks before it spends.
| It can | It asks first when |
|---|---|
| Research something new and build a subject | The name could mean more than one thing |
| Calibrate and maintain profiles, judging borderline articles itself | Never |
| Backfill past coverage | The cost is above your auto-approve limit |
| Work out which subjects a question is about | Never |
| Plan the judgements, estimate the cost and find a cheaper plan | Never |
| Run the analysis, sampled first and then in full | The cost is above your auto-approve limit |
| Group, chart and explain results, with citations | Never |
| Pin cards and arrange dashboards | Never |
| Create and change monitors | Always: monitors notify people and cost credits each month |
Sources
English-language web news, read with care.
Typestream finds articles through publishers’ RSS and Atom feeds, Google News sitemaps and GDELT’s public news feed. Its crawler follows robots.txt and keeps a polite pace. Paywalled articles are judged on their headline and standfirst, and marked as partial.
You see a headline, a short snippet and a link to the publisher, never the full text. How our crawler works, and how publishers opt out.
See what the coverage says about you.
Start a 30-day trial with 2,500 credits, and watch your first subject in minutes.