Documentation

In Claude, open Settings → Connectors → Add custom connector and enter:

https://verify.noevant.ai/mcp

Claude will send you here to sign in and approve the connection. Once you approve, the connector is live and Claude can call it. Custom connectors work on Free, Pro, Max, Team, and Enterprise plans; Free is limited to one custom connector.

Turn on Developer mode first — Settings, then Apps (Connectors on older builds), Advanced settings. A Create button appears once it is on. Enter https://verify.noevant.ai/mcp with OAuth authentication, approve, then enable it in a chat from the connected apps menu. Needs a paid plan; Free does not support custom connectors.

From the CLI:

openclaw mcp add fidenta --url https://verify.noevant.ai/mcp
openclaw mcp login fidenta

Or in the Control UI under MCP, or inline from the composer with + → Connectors → Add MCP server…. If tools do not appear, run openclaw mcp doctor --probe.

Perplexity supports the same MCP endpoint. Add https://verify.noevant.ai/mcp as a custom connector and authenticate the same way.

Not inside the research run itself — deep-research modes only call tools named search and fetch, which is a different job from returning a verdict. Fidenta works the other way round: run the research, then paste the finished report and ask your assistant to verify it. You get a verdict per claim instead of an overall impression, so the two sentences that are wrong are visible rather than averaged into prose that reads well.

No. If you do not have one, the sign-in step will offer to create one. New accounts start with 40 verifications and then 10 per month, with no card required.

Yes. Existing tokens keep working — sign-in was added alongside them, not in place of them. If your client supports a request header, send Authorization: Bearer <your token>. Get or rotate a token from your account page.

Never put a token in a URL query string. It ends up in server logs, proxy logs and browser history, and the MCP specification prohibits it.

It takes a claim, searches the live web for evidence, and returns a verdict with the sources it used. Two models from different labs are involved: one assesses the claim strictly against what was retrieved, and a second from a different lab audits whether that verdict is actually carried by the evidence.

The second pass exists because a model grading its own answer tends to agree with itself.

Supported — the evidence backs the claim.
Contradicted — the evidence says otherwise.
Unsupported — we searched and found nothing either way.
Unverifiable — it is an opinion or a prediction, so no amount of searching settles it.

Unsupported is not the same as false. Keeping those two apart is the whole design. Collapsing them is how a checker starts inventing failures that were never there.

When a claim is broadly right but the evidence carries an important qualifier, the result includes a caveat. A liability cap that has carve-outs elsewhere in the same contract is accurate as stated and materially misleading without the qualifier — the caveat is where that gets said.

Usually 10–25 seconds, depending on how many claims are in the text and how much searching each needs. If a check runs past your client's timeout you will get a "still verifying" response — ask again shortly and the finished result comes back. You are not charged twice for that.

Yes, and asking again for a claim already checked returns the saved result without spending another verification. Pass force: true to make it re-check against current sources — useful when the underlying facts may have moved.

Fact-checks text against live web evidence. This is the tool that spends a verification.

text (required) — the text to check.
force (optional) — re-check even if a saved result exists.
max_claims (optional) — cap how many claims to check, and so cap the cost.
confirm (optional) — required once the text contains more than three claims.

Looks up whether a claim has already been verified on your account. Free — it does not spend a verification.

claim (required) — the claim to look up.

Breaks text into checkable claims and reports how many there are and what verifying them would cost. Verifies nothing and charges nothing.

text (required).

Buys more credits in the conversation. Called without confirm it quotes the price and the card on file; it only charges on a second call after you have agreed. If no card is saved it returns a secure checkout link rather than charging.

Reports how many verifications remain in the current period. Free. No parameters.

Free accounts get 40 verifications on signup, then 10 per month. When you need more, 100 additional verifications are $19 and they do not expire.

One credit checks one claim. Paste a paragraph with six claims in it and that is six credits; a single sentence is one. estimate_claims, recall_claim and check_usage are free.

A claim already checked in the last 14 days is free to return. If you verify the same figure across three drafts, you pay once. Past 14 days it is charged again — not as a penalty, but because a saved verdict that old is no longer evidence of anything current, so the whole check genuinely runs again. estimate_claims tells you how many of your claims are already on file before you spend anything.

Anything that does not come back — a claim still searching when your assistant stops waiting — is refunded automatically.

Yes. For anything longer than a sentence or two your assistant checks first: it will tell you how many checkable claims it found and what that will cost, and wait for you to agree before anything is charged. You can also cap it — "check the first five" — to bound the spend on a long document.

Up to 20. That is enough for a news article or a section of a contract. Longer text is broken into claims and the first 20 are checked.

Your monthly allowance is drawn down first, then purchased credits. That ordering is deliberate: purchased credits never expire, so spending the perishable balance first is what you would want.

Yes, after your first purchase. Card details have to be entered on Stripe’s secure checkout the first time — they never touch our servers. If you opt in there or on your account page, later top-ups are a single confirmation in the conversation.

It is off unless you switch it on, you set a ceiling on how much can be charged that way in a month, and every charge is receipted. Remove the card in the billing portal to stop it entirely.

Calls return a clear quota message with a top-up link rather than failing obscurely. Nothing is lost — your history and saved results stay put.

Inference runs on models hosted inside our own infrastructure. Your text is not sent to an external AI vendor's API to be analysed. Retrieval reaches public web sources, which necessarily carries the search terms needed to find evidence.

No. We do not use your verification content to train, fine-tune, or evaluate any model, and we do not permit our service providers to do so.

The rating is stored against your own record so you can see it later. Separately we increment counters recording only which verdict type was rated and why — never the text of what you checked. That tells us where the verifier is weakest without anyone reading your content.

365 days from your last interaction with a given verification. Re-checking or rating one starts that period again. Sign-in codes are discarded after 10 minutes or on first use.

Email privacy@noevant.ai and we will remove it. Self-serve deletion is on the roadmap.

The privacy policy is at noevant.ai/privacy-policy. If you need a data processing agreement for business use, contact info@noevant.ai.

Check the URL is exactly https://verify.noevant.ai/mcp — including the /mcp path. A missing path is the most common cause.

Either the connection was never approved, or the token is wrong. Remove the connector and add it again to re-run sign-in. If you are using a token, confirm it is sent as Authorization: Bearer <token> with no Bearer prefix typed into the token field itself — that doubles it.

Tokens are stored hashed and cannot be shown again. Sign in at verify.noevant.ai/login and rotate it — the old one stops working immediately.

Open your verifications and mark it "Not right" with a reason. Those ratings are how we find where the verifier is weakest, and they are read as aggregate counts, never as your claim text.

Email info@noevant.ai.