Clear answers.
Honest limits.

Vigil is a control and verification layer for AI agents: short-lived signed permissions that your gateway enforces, tamper-evident receipts for the records you submit, watchdogs that alert on missing check-ins, storage for context you encrypt yourself, and paid source checks that an agent can buy per request.

Developers and small teams running agents that can touch something real: a payment provider, a production database, a customer inbox, a deploy pipeline. It also serves autonomous agents that need to check a source before relying on a claim.

If your agent only drafts text and a person reviews everything before it takes effect, you may not need Vigil yet.

An agent with credentials can act faster than anyone can watch. Vigil gives you a permission you can withdraw, a record you can verify later, an alarm for silence, and a bounded answer to “does this source support this claim.” Each piece states plainly what it does and what it does not guarantee.

No. Your agent runs wherever it runs today. Vigil issues permissions, signs records, and answers verification requests over HTTPS. It never executes actions on your behalf and never holds your provider credentials.

Vigil is a plain HTTPS API plus a Model Context Protocol (MCP) server. If your stack can send an HTTPS request or connect an MCP client over Streamable HTTP, you can integrate it. The examples use curl, Python, and TypeScript.

There are no framework-specific plugins yet, and not every framework, MCP client, or wallet has been tested. If yours behaves unexpectedly, email the details and it will be looked at.

Yes. Free includes one guarded agent, ten receipts per day, one watchdog, four journal slots, and REST and MCP access, with no card. It is a real tier for real work, not a time-limited trial.

Not today. Vigil is a hosted service; its value as a witness comes from being run separately from the systems it records. The verification side is open by design: public keys, exported chains, leases, and receipts can be checked with any Ed25519 library, offline, without Vigil’s software.

Vigil is built and run by its founder, Kenan Duncan, as an independent business. There is no outsourced support desk and no sales team. The person who answers your email is the person who wrote the code.

They keep working. Receipts and leases are signed documents that verify offline against Vigil’s published public key, and the chain can be exported at any time. Nothing you have exported depends on Vigil staying online. If the service were ever discontinued, the Terms require at least 30 days’ notice and a pro-rated refund of any unused subscription time.

A lease is a short-lived, Ed25519-signed permission. It names an agent, an action, an optional policy (the lim field), an expiry, and a signed recheck deadline. Your gateway checks all of those before it forwards an action to your provider.

A lease lasts from 5 seconds to 7 days, including renewals.

Revoke the lease. Your gateway learns about the revocation when it renews the lease or re-checks its status, and at the latest by the signed recheck deadline, which is at most 30 seconds after the lease was granted or last renewed. From then on the gateway denies the next gated action.

The stop happens in your gateway, not inside your provider. Vigil issues and withdraws the permission; your gateway enforces it.

Within 30 seconds at most. A gateway that checks online can see the revocation sooner; one that verifies offline treats the lease as invalid once the recheck deadline passes. Revocation is never instant, and Vigil does not describe it that way.

No. Your gateway holds them. Vigil never sits between your agent and your provider, so Vigil cannot act on your provider account, and a Vigil outage stops gated actions rather than allowing them.

Gated actions stop. A lease your gateway cannot renew or re-check becomes invalid at its recheck deadline, at most 30 seconds after the last grant or renewal, and a correctly built gateway fails closed. Actions that do not pass through your gateway are unaffected, because Vigil is not in their path.

Decide per action which side of that boundary it belongs on.

No. A refund that went through, a row that was deleted, an email that was sent: those are done. Vigil bounds what happens next and gives you a signed record of what was requested. Reversal, where it exists, is a feature of your provider.

A policy string that your gateway parses and enforces, such as amount<=50usd. Vigil signs it so it cannot be altered in transit, but Vigil does not interpret it and your provider never sees it. If your gateway ignores lim, no limit applies.

Yes, and it is usually small. Verify the lease signature against Vigil’s public key, check the agent, action, policy, expiry, and recheck deadline, then forward or deny. The developer guide lists each check.

A reference gateway in Python, built around Stripe refunds, is available on request while it is prepared for publication.

An active lease. Free allows one active lease at a time, Dev 25, Pro 250. Expired and revoked leases free their slot. One physical agent holding two active leases counts as two.

Yes. A lease is a signed document, so any Ed25519 library can check it against Vigil’s published public key. Offline verification is valid until the recheck deadline. After that, renew or re-check online.

A signed statement that Vigil saw a specific SHA-256 digest, under your account handle, at a specific time, linked by hash to the previous receipt. The signature covers all of those fields. You keep the receipt; the chain keeps the link.

No. You hash the record on your side and send only the 64-character digest and an optional public label of up to 200 characters. Vigil never receives the record itself. Keep private details out of the label.

Yes. Any receipt can be fetched by serial number without authentication, and the exported chain includes every receipt. A receipt shows the digest, the label, your account handle, the timestamp, and the signature. Choose a handle you are comfortable seeing next to your records.

It proves that this digest was submitted by this account and signed by Vigil at this time, and that the receipt has not been altered since. It does not prove that the underlying action happened, that the data was correct, or who wrote it. Treat it as signed-data integrity, not as testimony.

Fetch Vigil’s public key once through a channel you trust and pin it. Check the Ed25519 signature over the receipt’s signed fields. For chain integrity, export a range and recompute each receipt hash from the previous one. No Vigil software is required.

Not without detection, provided you pin the chain head. Every receipt hashes the one before it, so a deletion or reorder breaks the chain for anyone holding an earlier head. Deleting the newest receipts is the one change a fresh reader cannot see, which is why the docs ask you to pin the head. Vigil describes receipts as tamper-evident, not immutable.

Every six hours Vigil submits the current chain head to public OpenTimestamps calendars. Pro accounts can submit on demand and export the raw calendar responses. Those are pending submissions: Vigil has not verified inclusion in a Bitcoin block and does not describe any submission as confirmed. Independent confirmation needs the OpenTimestamps tooling and a Bitcoin data source on your side.

Every receipt and lease carries the public key that signed it, so older receipts still verify after a rotation. Keep every key you have trusted historically, and confirm a new key through an independent channel before trusting it.

Free: 10 per day. Dev: 5,000 per day. Pro: 100,000 per day. There is no cap on the total history you build over time.

A dead-man switch. You set an expected heartbeat interval, from 60 seconds to 7 days, and a public HTTPS webhook. Your agent sends heartbeats while healthy. If a heartbeat is missed, Vigil sends a signed alert to your webhook.

Vigil retries on every check for the first five attempts, then keeps trying once an hour. Alerts are never dropped silently. The next heartbeat resets the watchdog.

No. It tells you, or whatever listens at your webhook, that the expected signal is missing. What happens next is up to you.

Small private storage slots, up to 64 KiB each, for context your agent needs to keep between runs. Vigil stores the bytes you upload and returns them on request. It does not read, index, or interpret them.

No. Encrypt on your side before uploading, and keep the key. Vigil stores whatever bytes it receives and does not check whether they are ciphertext. If you upload plaintext, it is stored as plaintext.

Free 4, Dev 100, Pro 1,000. On a downgrade, the least recently updated slots beyond the new limit are deleted, so export anything you need before cancelling.

You send one public HTTPS URL and one claim of up to 2,000 characters. Vigil fetches the page in an isolated service, runs a bounded semantic assessment, and returns a verdict, exact quotations with character offsets, hashes of what it read, risk flags, and a signed receipt. The price is $0.03 USDC per check.

• supports: the retrieved text supports the claim.
• contradicts: the text contains evidence against the claim.
• mixed: the evidence points both ways.
• insufficient_evidence: the text does not settle it.
• not_assessed: no semantic assessment was made, for example because the page had no usable text.

Every verdict is about the supplied source only.

No. It means that one page, as retrieved at that moment, supports the wording you sent. An inaccurate page can support an inaccurate claim. Vigil does not verify the publisher, cross-check other sources, or rate credibility. Your agent decides whether the evidence is enough for its next step.

Public HTTPS pages on port 443 that serve HTML or plain text. Not supported: pages behind a login, content that only appears after JavaScript runs, PDFs, private or local network addresses, and following links to other pages.

The assessment covers the first 20,000 characters of the extracted text; longer pages are truncated and the result says so.

One bounded call to an Anthropic Claude model, with no browsing tools and no payment authority. The exact model version and methodology identifier are published in the live capabilities data and included in every result, so you can tell which version produced an answer.

Through x402, a payment protocol for HTTP. Your first request returns a 402 response with the exact terms: 0.03 USDC, the recipient, the network, and an expiry. Your payment-aware client authorizes the payment and retries the same request with the payment signature. Vigil submits the payment for settlement only when a result is released.

No. Source checks are pay-per-request and separate from Free, Dev, and Pro. A subscription does not include free checks, and a check does not require an account.

USDC on Base mainnet and Solana mainnet. Completed purchases so far have settled on Base; Solana is offered but has not yet been exercised with a funded purchase.

Any x402-compatible client should work; the published examples use a Coinbase-managed wallet with the x402 fetch wrapper, and the browser test page uses MetaMask on Base. Not every wallet has been tested.

Anyone with an x402-capable client can. There is no card checkout for source checks. The browser test page lets you try one purchase with a MetaMask wallet on Base; for automation, use the REST or MCP examples.

If Vigil cannot fetch or process the source (unreachable, blocked, oversized, unsupported type, timeout), the request fails with a stable error code and the payment is not submitted for settlement. If settlement fails, no result is released.

A completed result is charged even when its verdict is insufficient_evidence or not_assessed: you are paying for the bounded assessment, not for a particular answer.

Keep the idempotency key you sent and the payment proof. Retry the identical request with the same key, or fetch the private result with the recovery token from the response header. Vigil keeps results for 24 hours for this purpose. An uncertain payment is never a reason to authorize a second one.

No. The public receipt for a check carries a generic label and a digest; it names neither the URL nor the claim. The full result, including the claim and the retrieved text, is kept privately for 24 hours so it can be recovered, then pruned.

Yes, with the vigil_verify_source tool at the same MCP endpoint. Payment still requires a payment-aware wallet integration. An ordinary MCP connection cannot authorize a payment, and tool text is never treated as authorization.

Choose a handle and Vigil returns an API key that starts with vigil_sk_. That is one request, or one click on the pricing section. No email, password, card, or verification step.

A key is the only credential an agent needs, and it keeps Vigil from holding your email at all. The trade-off is yours to manage: the key is shown once, and there is no reset. Save it in a password manager or secret store the moment it appears.

No. Vigil stores only a hash of your key and holds no email or identity to verify you against. Register a new handle and move your integrations to it.

If the lost key had a paid plan, cancel it yourself through the Stripe customer portal using your checkout email, or email hello@vigilnotary.com with your Stripe receipt.

There is no in-place rotation yet. Register a new handle, switch your integrations, then revoke the old key. Revoking a key ends its authentication, revokes its active leases, and removes its watchdogs.

Delete journal slots before revoking; revocation does not erase them. Public receipts remain in the chain by design.

Yes. Each handle is a separate account with its own key, plan, and limits. Registration is rate-limited per IP address.

One key per account. Teams share it through a secret manager today. Separate seats and roles are not available yet.

Your plan, this month’s usage against each limit, the number of receipts and calendar submissions on your record, when your record started, and the public key to verify against. It reads the key from your browser’s local storage or asks you to paste it, and sends it nowhere except to Vigil.

Pro also includes on-demand calendar submissions with raw proof export. Source checks are billed separately at $0.03 USDC each.

Pick Dev or Pro on the pricing section. If your browser already holds a key, checkout attaches the plan to that account; otherwise you create a key first. Payment is a monthly card subscription through Stripe. The same key gains the new limits as soon as the payment is confirmed.

Yes. Start a Pro checkout with your existing key and the account gains Pro limits when the payment is confirmed. Each plan is its own Stripe subscription, so email hello@vigilnotary.com afterwards and the Dev subscription is cancelled with its unused days refunded.

Open the Stripe customer portal with the email you used at checkout and select cancel, or email hello@vigilnotary.com. Cancellation takes effect at the end of the month you have paid for. Your key and receipts stay; the account returns to Free and capacity above Free limits is removed then.

Your key and your receipts stay. Capacity above the new plan is enforced immediately: the newest watchdogs beyond the limit are deleted, the newest active leases beyond the limit are revoked, and the least recently updated journal slots beyond the limit are deleted. Export what you need first.

Yes, in two cases without questions: the first charge on a new subscription, refunded in full if you ask within 14 days, and a renewal you meant to cancel, refunded in full if you ask within 7 days and did not use paid capacity after it. If Vigil discontinues a paid feature or ends your subscription without cause, the unused time is refunded pro rata. Other partial months are not refunded. A refund returns the account to Free.

Source checks paid in USDC are never charged when the check fails and are final once a result is released.

Cards through Stripe for Dev and Pro. USDC on Base or Solana for source checks. No invoicing or bank transfer yet.

Prices are in US dollars. Any applicable tax is shown at Stripe checkout before you pay. Stripe sends a receipt for each payment; email if you need an invoice with specific company details.

Not yet. If you need more than Pro provides, email hello@vigilnotary.com and describe the workload. Vigil does not claim enterprise readiness.

The current prices are the ones shown on this site. A subscription price increase is announced at least 30 days ahead on the site and by email to the address on file with Stripe, and applies from your next renewal after that date. Agents should read the live payment terms before every purchase rather than caching a price, because the 402 response is the authoritative quote for that request.

Your handle, a SHA-256 hash of your key, your plan, and Stripe customer and subscription identifiers if you pay. The receipts you create, your active and past leases, watchdog names, intervals, and webhook URLs, and your journal bytes. For source checks, the request and result for 24 hours.

No email address, no name, no card details: Stripe holds those.

No. The site loads its own scripts and fonts from vigilnotary.com only. There is no analytics or advertising code, and no cookies. The server keeps standard access logs; payment signatures, idempotency keys, and recovery tokens are stripped from them.

On Hetzner Cloud, behind Caddy with automatic TLS. Nightly database backups are taken. Vigil is currently a single-region service. Every provider that touches your data is named in the Privacy Policy.

No third-party audit has been performed yet, and Vigil does not claim one. The code has been through internal adversarial review and carries an automated test suite that includes payment-failure and tamper cases. Report a vulnerability to security@vigilnotary.com; good-faith research is welcome under the disclosure terms.

Not yet. Vigil is a public pilot and does not offer a service-level agreement. Build your gateway to fail closed and your watchdogs to alert independently, so a Vigil outage stops gated actions rather than hiding problems.

The Ed25519 private key lives only in the production server’s configuration, never in the code repository or the browser, and the service refuses to start in production without it. The public half is published and embedded in every receipt and lease.

No. Despite the domain name, Vigil is not a notary public and does not perform notarial acts. A Vigil receipt is a cryptographic attestation that a digest was signed at a time. Whether that is useful evidence in a dispute depends on the jurisdiction and on how you handle the underlying record. Ask a lawyer for that part.

Vigil holds no certifications and claims no compliance framework. It stores little personal data by design, names every sub-processor, and honors access, export, and deletion requests for everyone. To delete your account, delete your journal slots, then revoke your key. Receipts are public by design and cannot be removed from the chain; they contain your handle, a digest, and a label only.

No. It bounds actions that pass through a gateway you control and gives you evidence about the rest. An agent with its own credentials, an action that bypasses the gateway, or a mistake inside a permitted boundary is outside Vigil’s reach. That is why the limits are stated on every page.

Email hello@vigilnotary.com. Dev accounts get email support and Pro accounts get priority; Free accounts get answers as time allows. There is no phone support and no live chat.

Not yet. Free-tier “community support” means the documentation, the public examples, and email replies as time allows.

Yes, by email. Say what your agent can touch and what you wish Vigil did about it. The roadmap follows what people put behind a lease.

Three pages, each with a plain summary at the top of every section: the Terms of Service, the Privacy Policy, and the Refunds and cancellation page. Checkout links to all three, and a separate step asks for your consent to the monthly renewal.

The agent reference describes the running service in plain language, and the developer guide has copy-and-paste examples for every capability.