Evaluator builds arguments from sources you trust. It searches only the websites and local files on your list, and every point in a result links back to the page or passage it came from.

1. Get two free keys: a Google Gemini key and a Tavily key. How to get them.
2. Open Settings › AI model and search. Choose Google Gemini, paste the Gemini key, click Load models and pick a Flash model. Paste the Tavily key under Search.
3. Look through Trusted sources and Topics. A starter list is already there; switch off anything you don't trust.
4. Go to Trace, choose a Topic, type a statement, and press one of the four buttons.

Button	What it does	Good for
Support	Builds the strongest case for the statement.	Finding evidence for a thesis
Challenge	Builds the strongest case against it.	Testing your own argument
Evaluate	Builds both cases and weighs them in one balanced report.	Contested or interpretive questions
Verify	Checks whether the statement is correct, claim by claim. Logic and arithmetic are checked exactly by the app.	Scientific, mathematical and logical statements

Results are saved on this device under Recent results, and can be copied, printed or saved as PDF, or downloaded for Word.

What you need

• A recent browser: Chrome, Edge, Firefox or Safari, updated within the last couple of years, on a computer, tablet or phone.
• An internet connection to run analyses. Past results and your settings can be read offline.
• Site storage allowed. The app keeps your keys, sources and results in the browser. Private or incognito windows may block this, or forget everything when closed.
• For local files: any of the browsers above can read files you select. Only Chrome and Edge on a computer can remember a folder so you can rescan it later.
• To install it as an app: the site must be served over HTTPS. In Chrome or Edge use the browser menu › Install app; on iPhone or iPad use Safari › Share › Add to Home Screen.

How the pieces fit

Every analysis uses two services: a search that is restricted to your trusted websites, and an AI model that reads what the search found and writes the result. You bring your own keys, which are stored only in this browser and sent only to the service they belong to.

Choose an AI model provider

Free tiers have daily or monthly limits. If you hit one, switch provider in Settings; the app remembers a key for each.

Set up search

• Tavily works with every provider. Get a free key at app.tavily.com: 1,000 searches a month, no card needed.
• The model's own web search is offered only for Claude and OpenAI, the two whose search can be limited to your trusted websites. It's slower and billed by them, but needs no Tavily key.
• If a topic contains only local folders, no web search or Tavily key is needed at all.

Put the keys in

1. Open Settings › AI model and search.
2. Choose a Provider and paste its key.
3. Click Load models and pick one. Optionally pick a Fast model for planning searches, such as a "Flash-Lite", "Haiku" or small Groq model.
4. Choose the Search and paste the Tavily key if you use it.
5. Check Balance and usage. OpenRouter and Tavily show what you have left; the others link to their billing pages.

Settings save as you change them. The Trace page shows which model and search are in use, and names anything still missing.

Your trusted list is the heart of the app: searches never go beyond it. Topics group sources by subject, so a question about Assam can search only the archives and journals you trust for Assam.

Add a trusted source

1. Open Settings › Trusted sources › Add a source.
2. Domain: type the site's address without https://, for example archive.org. Add a path to trust only part of a site, such as loc.gov/collections. Subdomains are included automatically.
3. Name, Type (primary sources, academic journal, reference work, and so on) and Notes help you and the model judge the evidence.
4. Tick the Topics it belongs to, then Add source.

Use the switch beside a source to leave it out of searches without deleting it. Edit changes any detail; the Show menu filters the list by topic.

Create a topic and map sources to it

1. Open Topics › Add a topic. Give it a name and, optionally, a description.
2. Tick the sources (and local folders) that belong to it, then Add topic.
3. To change it later, open the topic and press Edit. Deleting a topic keeps its sources.

You can map from either side: tick sources inside a topic, or tick topics when adding or editing a source. A source can belong to several topics.

Use a topic

On the Trace page, pick it from the Topic menu. Only sources that are both in that topic and switched on are searched. All trusted sources searches everything that's switched on.

Back up or share your list

• Export JSON saves your sources and topics to a file; Import JSON replaces them with a file's contents. Use this to move your list to another device or share it with a class.
• Restore default list brings back the starter sources and topics (and replaces yours).
• Claude's own search accepts up to 64 websites at once and OpenAI's up to 100. With larger lists, choose a topic or use Tavily.

Your own documents (lecture notes, articles, reports, theses) can be sources alongside websites. Their text is read in your browser and stored only on this device.

What can be read

• Yes: PDF files that contain text, Word files (.docx), plain text (.txt) and Markdown (.md), up to 50 MB each.
• No: scanned PDFs (pictures of pages, with no selectable text) and old Word files (.doc). Save a .doc as .docx or PDF first. Other file types in a folder are skipped.
• Files that can't be used are listed with the reason, so you know what was left out.

Add a folder or files

1. Open Settings › Local files.
2. Add a folder reads every supported file in a folder and its subfolders. Add files reads just the files you select.
3. Wait for the progress bar to finish. Reading large PDFs can take a little while; it happens only once.
4. Open the folder's entry and tick the Topics it belongs to. If a topic was selected on the Trace page when you added it, it's already mapped there.

Keeping it up to date

• In Chrome or Edge on a computer, press Rescan after adding or changing files.
• In other browsers, Remove the folder and add it again.
• Search this folder switches it on or off; Rename changes its display name. Removing it deletes only the stored text, never your files.

In results

Each file gets its own colour, and citations show the file and page, for example thermodynamics.pdf, page 1. Open Show passage in the source list to read the exact text used.

Privacy: your files never leave the device as a whole, but the passages chosen as evidence are sent to your AI provider, just as web excerpts are. Don't add documents you aren't allowed to share with that provider.

• Colour key: each website or local file has its own colour. A tinted sentence was drawn from that source, and the number after it opens the source in the list at the bottom. Text without colour is the model's own connecting reasoning.
• Removed citations: if the model cites a page the search never returned, or one outside your trusted list, the app removes it and says so in the footnote.
• Summary / Detailed switches between the short view and the full evidence.
• Verdicts: Support and Challenge rate the strength of their evidence; Evaluate gives an overall leaning; Verify grades each claim from True to False, or Unverified when your sources don't cover it. Unverified is not the same as false.
• Exact checks in Verify (logical validity and calculations) are worked out by the app itself, not guessed by the model.
• Reused evidence: a repeat run on the same statement reuses the earlier search for up to 6 hours. Press Search again for fresh results.
• Search focus: in Settings, choose whether Support and Challenge search only for their own side (more targeted) or share one search (faster, fewer credits).
• The footnote records the topic, sources searched, model, tokens used and how long it took.

Treat results as a research assistant's draft

• Open the sources. Citations are checked against what the search returned, but a model can still misread or overstate a passage. Before relying on a point, read the source it cites.
• Cite the original sources, not this app, in your own work, and follow your institution's rules on disclosing AI use.
• A strong case isn't proof. Support and Challenge are advocates by design; each argues one side. Use Evaluate or Verify when you need a judgement.
• For important questions, cross-check: run it again with another model, or switch search focus. Agreement across runs is more convincing than one answer.

Get better answers

• Be specific. "The Treaty of Yandabo (1826) ended Burmese rule in Assam" works better than "Burma and Assam".
• One claim at a time gives clearer results, especially in Verify.
• Choose the right topic. A focused list of good sources beats a long list of mixed ones.
• Read "Gaps in the sources". It tells you what your list couldn't answer, which is often where more reading is needed.

Mind the limits

• Your list shapes the answer. If it leans towards one school of thought, so will the results. Include sources from different perspectives.
• Paywalls (JSTOR, many journals) often mean only abstracts are searchable.
• Interpretive questions ("Was this policy a failure?") have no single true answer. Use Evaluate, not Verify.
• Verify's truth-table check covers arguments of the form "if…then", "and", "or", "not". Arguments using "all" or "some" are judged by the model, and the result says so.

Privacy, keys and cost

• Don't include personal or confidential information in statements or local files; it's sent to the AI provider.
• Keep your API keys private, like passwords. Remove them from shared or public computers (Settings › Remove this key).
• Set a spending limit in your provider's billing settings when using paid models, and watch Balance and usage.

"The API key was rejected"

Check it was copied in full, with no spaces, and that billing is set up for paid providers. Some keys take a minute to start working.

"You've hit a rate or usage limit"

Free tiers limit requests per minute and per day. Wait a minute, lower Searches per trace, or switch provider.

"The browser couldn't reach …"

Check your connection. If it keeps happening with one provider, that provider may not accept requests straight from a browser; choose another.

Nothing was found

The topic's sources may not cover the statement. Rephrase it more specifically, choose a broader topic, or add sources.

"The model's answer couldn't be read"

Small free models sometimes produce malformed answers. Run it again, or pick a larger model.

Local files won't add

Private windows often block on-device storage. Use a normal window, and check the file isn't a scanned PDF or an old .doc.

An update doesn't appear

Reload the page twice: the first reload fetches the new version, the second shows it.