Research
Plan, read, verify every quote, write a report you can audit. The one screen that always needs an LLM key — and the only research tool that shows you its evidence.
On this page
Compose
- Open Research. Type the question — up to 2,000 characters. Ctrl/Cmd+Enter runs it.
- Pick Effort. Each segment’s label carries its own projection:
| Effort | Sub-researchers | Tool calls |
|---|---|---|
| quick | 1 | 10 |
| standard | 3 | 45 |
| deep | 5 | 100 |
- Press Run. The cap beside the button is the guarantee; the projections are estimates and say so.
The Advanced drawer shapes the run: Allow, Deny and Prefer domain lists; Read with my login (saved sessions, acknowledged once per domain — read-only, one page every 10 seconds per site); From/To publication dates; Search backends (none picked means the keyless ones); Writer and Judge provider/model pairs; the Cap $; and Review the plan first, which pauses the run after planning for approval or a redraft.
The run panel
Live phases show as a pill — planning, review, running, writing — over a budget meter and two counters: facts so far and usable facts. While it runs:
- Steer — type a message; the run reads the latest one at its next step.
- Write now — skip to the report from what is already verified. Disabled until a plan exists.
- Stop — keeps the checkpoint. Resume continues the same run, optionally with a higher cap.
The activity log streams every step. The couldn’t-read list names each page that defeated it and why: login-required, access-denied, unavailable, empty, challenge, robots, error.
The report
The finished report renders with citation chips, each toned by its fact’s verdict — verified, softened, contested. Click a chip: the evidence drawer shows the claim, the exact quote, the passage around it, the URL and the checked-at timestamp. The quote was substring-matched against the stored snapshot before any model judged it; URLs the writer invents are stripped.
Below the report: the verification panel, the findings table (the whole ledger, dropped rows included — the audit trail), and the cited sources. Exports: Markdown, Sources CSV (footnote, url, quote, checked_at, content_hash) and BibTeX. DOCX or PDF is a copied pandoc line.
Run again repeats the question with the same options and approved plan, then diffs the evidence against the earlier run: sources added and gone, pages changed, facts added, gone, and quotes whose verdict moved. Forget (on a recent run) deletes the run, every stored page copy and its read trail; its spend stays in the budget history.
Budgets and caps
The cap defaults to Settings’ max_cost, else $1.00. Research stops researching at 85% of it so the report still gets written. The run’s spend counts against the monthly budget like any metered call. The CLI equivalent — magpie research "<question>" — prints an estimate to stderr before any spend (CLI reference).