NewAnswers MCP: search and quote-checked answers for your agents.See how it works
You get the report and everything behind it: every source it read, the passage each citation rests on, and a verdict on each one. Export it as plain Markdown, or let your own agent search it.
4,069 stars on GitHubOpen source, MITView the projectStore
Each source is saved with where it came from and when, the full text wherever a legal copy exists, and the exact quote behind each claim. A follow-up question in the same project reads from it first.
How the vault worksGet
Each run reads web pages in full, then calls more than 40 data APIs directly for official numbers and records: Census, BLS and BEA tables, SEC EDGAR filings, OpenAlex and PubMed papers, recalls, sanctions and city open data.
What a run can readValidate
Each citation gets a verdict on the receipt. Cited figures are checked against the source text and corrected where needed before the report ships, and retracted papers are flagged.
What the receipt showsFor agents
Answers MCP gives an agent two tools, and both return the exact words a fact rests on.
Search 5¢ a call · ask from 5¢ · $5 free credit, no card needed
answers.hyperresearch.ai/mcpEvaluation
DRACO benchmark (Perplexity × Harvard) · 10 tasks, same judge · October 2026How we tested →
DRACO grades a finished research report against expert-written criteria for factual accuracy, breadth, presentation and citations, and takes points off for wrong claims. We ran every system on the same tasks and graded each report with the paper’s grader and the same judge. Source counts are per-run averages on the first 3 tasks. Method and per-task scores →
Deliberate at every step
This is a Deep run. It plans the question, researches it chapter by chapter, writes the report in one pass, and checks every cited claim before it ships. A Light run takes the short path: one sweep, one draft, a final edit and the same claim check.
Break down the question, split it into research chapters, and gather a broad foundation of sources for each.
Your question broken into the things it actually asks · required headings · coverage matrix
3–5 research chapters, researched two at a time
Your vault first, then planned searches · breadth, depth, adversarial lenses · parallel fetcher waves
In every chapter: find contradictions, investigate the important questions, and challenge the source collection.
Opposing claims paired into ranked fight clusters
Two analysts pick where depth pays · scored, budgeted
One investigator per locus, in parallel · a committed position, not a summary
Committed positions reconciled into named tensions
Expert disagreements pulled from full bodies, not summaries
“What source would overturn this?” · gaps fetched before writing
Load-bearing claims and verbatim quotes, indexed
Map the evidence, plan every section, fetch what the plan still lacks, then write the whole report in one pass from all of the evidence.
Every kept source read in full · its facts, figures and quotes mapped
One plan over the whole map · framing, findings, a brief for each section
What the plan needs that the map lacks · a targeted fetch, primary sources first
One writer works from all of the evidence · the whole report in one pass
Check every cited claim against its source, correct what does not hold, and record what was checked.
Every cited figure read against the source it cites · one guarded revision
Per-citation verdicts, quote integrity, retractions · then the ship gate
What a run can reach today
Pay as you go · Runs from $9
Create an account and add a card: your first Deep run is free. What comes back is the report, the source list and a table of every claim with its quote, and a Deep report adds a verification receipt.
One platform, four possibilities
From the first question to the final citation, a connected toolkit for work that depends on getting it right.
Send one question at the size you need. Light is $9: about 30 to 40 minutes and an answer-first report of about 2,500 to 4,000 words, sized to the question, its cited claims checked against their sources before it ships. Deep is $49: 3 to 5 hours across about 150 to 200 sources, with a verification receipt saying which citations were checked and what the checker found. Either way you get a report that spans the field and argues a case inside it, plus the source list with where each page came from and how good it is, and a table of every claim with its quote. Every intermediate artifact is yours to download if you want to see the work.
Every source a run fetched, every interim note, every claim stays in your workspace. Search it by words or by meaning. Sources carry a quality score and are clustered so five copies of one press release count as one voice. The next run searches the vault before it touches the web, so your second question is cheaper and better informed than your first. Connect your own agent when you want to ask questions across the vault; answers stay in that agent and link back to evidence.
text-wrap CSS property - CSS | MDN
Your vault and your runs are exposed as a remote MCP server. Claude, Cursor, ChatGPT, or an agent you wrote can search notes, read sources, start runs, and check citations from inside the tool you are already working in. Your client opens a browser authorization screen where you choose the workspace and permissions. For your own pipelines, the same runs, vault and reports are on a REST API with scoped keys, and everything also works from the console in your browser. A run is long and asynchronous: start it, get its id back at once, and take the finished report when a webhook says it is ready. The OpenAPI document and the webhook guide are in the docs.
› search_notes {"query":"text-wrap balance","mode":"hybrid","limit":2} {"hits":[ {"note_id":"text-wrap-css-property-css-mdn","title":"text-wrap CSS property - CSS | MDN"}, {"note_id":"css-text-wrap-balance-css-and-ui-chrome-for-developers","title":"CSS text-wrap: balance | CSS and UI | Chrome for Developers"} ]}
Hand it a document and its sources. You get a verdict for each citation: supported, partially supported, unsupported, or wrong source. It also checks that every quoted passage is really in the source, sweeps for retracted papers, and flags sources that are secretly the same story. Run it before you publish.
The shorthand text-box: trim-both cap alphabetic eliminates font-metric whitespace in Safari 18.2 and Chrome 133. partially-supported
source saystext-box-trim is initially set to trim-both, meaning that the over and under edges of the paragraph are trimmed. text-box-edge is initially set to cap alphabetic, meaning that the text is trimmed flush with the top of capital letters…
suggested fixCross-reference [[css-text-box-trim-blog-chrome-for-developers]] for release provenance.
Confidence comes from the work
Good research should invite scrutiny. We make the evidence part of the answer.
When a paper sits behind a paywall, the service looks for a legal open-access copy and reads that whole, not the abstract.
A run reads across the field first, finds where the sources actually disagree, and only then sends investigators down those specific questions.
Before anything is written, a critic asks what source would overturn the direction the research is taking, and those gaps are fetched first.
Every cited claim is read against the source it cites before the report ships, and a retracted paper the report does not acknowledge stops it.
Everything fetched is fenced as untrusted data before a model sees it.
The research stages are the open-source CLI's. The hosted Deep run then plans the report and writes it in one pass.
Where a run can read
Every rung is recorded on the source it fetched. A question asked in another language is searched in that language as well as English, a Chinese question on Deep ranks official Chinese sources first, and the report can be written in English, Chinese, Japanese or Korean.
Every run searches what earlier runs already read before it touches the web.
A fetch tries a direct request first, then Cloudflare browser rendering for pages that only exist after JavaScript runs.
OpenAlex and PubMed find the papers, Crossref supplies the DOI record and any retraction notice, and SEC EDGAR returns filings, Form D notices and five-year financials.
For any DOI the pipeline finds, it asks Unpaywall, Europe PMC and CORE for a legal open-access copy and reads the whole paper.
When a question turns on what people said as something happened, the planner opens a search lane on X for that part of the question.
Connect your agent
Your runs and sources are a remote MCP server. Authorize it from your own client and ask questions across everything it has read.
Make your next question the start of something bigger.
Every source it reads stays in your vault, and the next question starts from there.
