A desktop research agent that plans its own questions, searches the live web, verifies every claim against its sources, and writes cited reports.
AI research tools invent sources. You get confident prose with citations that do not resolve, or resolve to pages that never said that — which makes the output unusable for anything that matters.
- Classifies each message as conversation, memory lookup or a research run, so ordinary chat costs nothing
- Plans 10+ questions, searches, ranks and reads sources, then verifies each claim against the evidence
- Streams back a cited report with a live panel showing progress, questions and sources
- Remembers across conversations, and answers with provenance
Fabricated citations are structurally impossible
The model only ever sees opaque IDs like [S1]; the backend resolves them to real stored sources. An unknown ID triggers a repair pass, and if that fails the turn fails rather than shipping an invented source.
Context that does not grow without bound
Each turn rebuilds the prompt from six prioritised blocks against a fixed token budget, and summarisation is incremental. Turn 200 costs the same as turn 20, which is what keeps running costs flat as usage grows.
Partial failure is a first-class state
A run that extracts 12 of 24 sources still produces a report, flagged as degraded with the reason recorded, instead of failing outright.
Retrieved content is treated as hostile
Fetched text is wrapped in an untrusted-data fence, and SSRF filtering runs in our own code before any fetch.
If you need research, monitoring or reporting that a person would otherwise do by hand — and the output has to be trustworthy.
Need something like this?
Tell us what you are trying to build. A few lines is enough to start, and you get a reply within one business day.