Legal

Privacy Policy

Last updated: 2026-10-28

Zero data plane architecture

Core Spend SDK policy evaluation runs in your application runtime. For supported provider bindings, prompt and completion content travels between your application and the selected model provider; AgentGuard is not an inference proxy. Customer-supplied provider credentials and locally configured signing private keys are not sent to AgentGuard by those bindings.

This boundary does not apply to data you intentionally submit to AgentGuard's hosted website, dashboard, account, support, billing, consent, workflow, or verification features.

What we DO receive or process through hosted features:

What we do NOT receive from governed Spend SDK calls:

Setup email and optional product updates

When you request setup steps, we store your email address, request date, delivery status and product-update preference in our hosted storage, and send the setup email through Postmark. The updates checkbox is optional and unchecked by default. Leaving it unchecked means you receive only the setup email.

If you opt in, we also send a setup check on day 3, titled "Did it install? Reply if anything broke", and a Team overview on day 14, titled "What Team adds". The automated sequence ends there. We may send occasional product updates manually to opted-in addresses. Each setup and follow-up email includes an unsubscribe link. Unsubscribing stops the remaining follow-ups and product updates, and we retain your preference to honor that choice.

We also use setup email domains to understand company interest. A daily internal digest sent to JP through Telegram groups successfully delivered setup requests by company domain, with the number of distinct addresses and the first-seen date. It excludes common personal email providers. Only domains, counts and dates are sent in this digest, without full email addresses or names. This aggregate use applies to setup requests whether or not you opt in to product updates.

Cookies and analytics

The dashboard uses cookies, local storage, or session tokens to keep you logged in and route you to the right workspace.

Analytics on this site is two things. Vercel Web Analytics is a cookieless visit counter: it runs on every page, sets no cookie and no identifier, stores nothing on your device, and reports the page, the referring site and coarse device and country data to Vercel. It counts clicks on a few named buttons and links the same way, for example Copy install link or Check pasted files: the button's name and nothing else, never anything you type or paste. PostHog does not run until you accept it in the cookie bar, or later through the Cookie settings link, and pages without the cookie bar never load it. Once you accept, PostHog records pageviews, scroll depth, the campaign tags a link carries you in on (for example utm_source and utm_campaign), and a session recording. The session recording masks every input field and the text of anything marked private, includes cross-origin iframe content on pages that have any, and samples about 1 in 10 sessions rather than every visit. If you accept, X's ads pixel also loads and reports the page view to X, so we can see which of our ads on X brought visitors; when you arrive from an X ad it sets a cookie named _twclid for up to 30 days, and X may set cookies on its own domain under its own policy. If you reject or never answer, PostHog and the X pixel never run in your browser; only the cookieless counter does. Separately, when you check a public repo with the harness check or open a shared harness result, run the template score or run the payment risk check (on the website or through the hosted MCP tools), or send an AgentGuard Score from the plugin or the AgentGuard Score page, our server records one count in PostHog: which tool ran (and, for an AgentGuard Score, whether the plugin or the page sent it), its result band or grade, and the day. That count carries no identifier for you and nothing you checked, and it is recorded whether or not you accepted analytics.

Partner and campaign links: if you arrive through one, its code rides along in the page address to checkout and to other pages on this site, so the partner is credited, whether or not you accept analytics. Nothing is stored for that by default. If you accept analytics, we also remember the campaign or partner link you arrived from for 30 days, in the storage listed below; rejecting removes it.

What this stores: once you accept, PostHog's own cookies and local storage keys, all named with a ph_ prefix, and, for 30 days, the campaign or partner link in keys named agentguard_first_utm and agentguard_builder_code, plus X's _twclid cookie when you came from an X ad. Rejecting removes all of these. Either way, your choice itself is kept in local storage under the key agentguard-consent, never in a cookie. Vercel Web Analytics stores nothing on your device, so there is nothing for this choice to remove; it counts as described above either way.

Change your choice any time with the Cookie settings link in the footer.

Third-party processors

Hosted features use processors visible in the local implementation, including Privy for authentication, Stripe for billing, Postmark for email, Vercel and Vercel KV or compatible storage for hosting and account state, PostHog and Vercel Web Analytics for analytics, OpenRouter for help-chat responses and managed inference, and optional Supabase, Hermes, or Telegram destinations. The core Spend SDK provider bindings do not route governed prompt or completion content through AgentGuard, but a selected model provider still receives the inference request.

Optional anonymous install count

Burn's anonymous count is off unless you say yes after a successful report or explicitly turn shared Burn and Spend telemetry on. The question is shown at most once on an interactive terminal. Enter, any answer other than y or yes, or no answer for 15 seconds leaves it off. Burn does not prompt or ping in CI, hooks, library use or a noninteractive terminal.

A yes allows one attempt per install to send burn_first_report to https://agentguard.run/api/beacon. Its exact fields are event, sdk_name (@agentguard-run/burn), sdk_version (Burn version), runtime (Node version), os_family, anonymous_install_id (a random ID shared with Spend), ci_env_detected (false) and tty (true). It sends no prompts, files, paths, project names, session IDs, model names or usage counts.

The choice and random ID stay in the local install.json shared with Spend. Run agentguard-burn telemetry status to inspect the choice, agentguard-burn telemetry on to enable it or agentguard-burn telemetry off to disable it. AGENTGUARD_TELEMETRY=1 explicitly enables shared telemetry. AGENTGUARD_NO_BEACON=1 always turns it off, including when the on switch is set.

Accepted Burn and Spend beacon events are counted as server custom events in our existing Vercel Web Analytics project. The durable count stores the event name, SDK name and SDK version, along with the server event time and a fixed endpoint address. Our Pro plan provides a 12-month reporting window, subject to Vercel's retention terms. The existing short-lived server log receives all eight sanitized fields and receipt time. Request IPs are used for temporary rate limits; we do not forward request IPs, cookies, user agents, install IDs or referrer URLs to these custom events. Opening Burn's Next link makes an ordinary website visit to /burn/next with no query string or install ID. That visit uses the cookieless page counter described above and needs no CLI ping.

Team efficiency and daily counts

Plugin 0.6.2 removed both uploads. Only versions 0.6.1 and earlier can send them, and agentguard.run has no endpoint for either, so nothing sent was stored. In those versions they exist only for Team licenses and are off by default. Free and Solo use sends nothing from them. Each one needs the person to turn it on locally, in the plugin folder, and each one can be turned off the same way. Both are sent by the plugin's background worker to agentguard.run with the Team license key as the bearer token, signed with the same local key that signs AgentGuard's receipts, after a session's closing receipt is written and never while a hook is waiting. Neither upload can carry document text, prompts, model output, file paths or client names: the payload is built from fixed fields only and anything else is rejected before sending.

Weekly waste causes (team efficiency)

Turned on with the environment variable AGENTGUARD_TEAM_WASTE=1 plus a sharing choice the person makes locally: "totals" or "seat". The worker then sends at most one previously viewed, completed UTC week per call to https://agentguard.run/api/team/waste. What is sent: the schema name, the Monday the week starts, a seat id (a SHA-256 hash of the license key and the machine fingerprint, so it is stable for one machine and never the person's name or email), and the week's Burn figures: tokens and list-price dollars (in micro-dollars) split into total, useful, waste and unattributed; the same two numbers for each waste cause (history re-sent, files re-read, sub-agent fan-out, reloads), for reloads by cause (idle, model switch, other), and for wrong-model use including what a smaller model would have cost; the sub-agent token total; the top waste cause as a code; and the public key and signature. Both sharing choices send these same per-seat figures, because team totals are added up from them. The choice decides what a manager can see: under "totals" the seat never appears on its own; under "seat" the manager view may show that seat's top waste cause and its suggested fix, and never that seat's shares, tokens or dollars. A separate signed request carries the sharing choice and the UTC time it was set. Turning sharing off sends a signed removal request for that seat's stored weeks at once, and the worker retries it if it fails.

Daily counts

Turned on with the environment variable AGENTGUARD_TEAM_COUNTS=1 plus the command node runtime/counts-sync.cjs on, run in the plugin folder; node runtime/counts-sync.cjs off stops it. The worker then sends one request per completed UTC day to https://agentguard.run/api/team/counts. What is sent: the date, the same hashed seat id, the day's totals of launches, allowed, asked, said yes, stopped, shadow, unrecorded and unresolved asks, the day's token total, and the public key and signature. A day is sent only when every session of that day has a complete verified receipt, at most once a minute.

The exact field lists live in the plugin's README under "What it sends, and where". If a field is not listed there, it is not sent.

Team seats and shared caps

A Team license counts one seat per person. Each person on a Team's roster, meaning the license owner and each invitation the owner sent, has a member ID that is available to the license owner. The plugin's background worker can add that member ID to its seat heartbeat, next to the license key, machine fingerprint, process id and loaded policy hash it already sends, so all of one person's sessions and machines count as one seat. We keep a hashed form of the member ID with the seat record, and the member ID itself with the roster entry and the email address the dashboard already holds for that invitation.

If a Team owner sets shared caps in the org policy, the worker reserves allowance from https://agentguard.run/api/team/reserve and returns what it did not use to https://agentguard.run/api/team/release. What is sent: the license key, the member ID or the machine's hashed seat id, whether the budget is tokens or US cents of priced tool calls, an amount, the current UTC day and month, and on release the grant id. Any other field is refused. No prompt, tool name, argument, file path, model output or session content is sent. We store, for each team and UTC month, the day and month totals and each open grant with a hashed reference to its holder, its day and its amounts. These records expire three days after their month ends. The IP address is used for rate limiting.

Data retention

Hosted account, license, billing, inquiry, support, analytics, workflow, and dashboard data is retained according to the feature and configured processor. The local source does not establish every processor's backup or deletion interval. A signed receipt that you retain remains verifiable with its public key after hosted data is removed; cryptographic signing does not itself publish or retain the receipt.

User rights

If you are covered by GDPR, CCPA, or similar privacy laws, you may request access, correction, export, or deletion of personal information we hold about your account. Email legal@agentguard.run.

International transfers

Hosted account, license, billing, inquiry, support, analytics, and workflow data may be processed in the United States or in regions used by the configured infrastructure providers. The governed SDK inference path remains application to selected provider and does not pass through AgentGuard infrastructure.

Children's privacy

AgentGuard is built for businesses and developers. It is not directed to children under 13.

Contact

Privacy questions can be sent to legal@agentguard.run.

Back to top

Functional-use disclaimer

All terminology and labels used in AgentGuard materials are descriptive of software functionality only, not legal definitions or guarantees of compliance. Terms such as receipt, audit log, evidence, audit trail, attestation, signed, verified, attested, compliance, compliant, outcome, settlement, capability tier, and settles refer to cryptographic software records and programmed state transitions only. They do not state that any record has binding legal effect, official certification, or equivalent status to records maintained by courts, banks, auditors, regulators, or agencies.

Terms drawn from audit, evidence, settlement, credit, debt, market, liquidity, maturity, and similar domains are used only in a functional and descriptive sense. An audit log is a sequence of recorded software events. Evidence means a record that may support a user review. Settlement means a final software state. None of these terms should be read as legal, financial, accounting, or regulatory advice.

Financial-context words such as trade, trading, liquidity, maturity, market, clearing, and exchange, if used in examples, describe token, budget, or workflow mechanics only. AgentGuard does not operate as a broker-dealer, exchange, clearinghouse, investment adviser, insurer, government agency, or regulated marketplace.

Words such as offer, obligation, credit, debt, payment, settle, maker, and taker refer only to hypothetical or user-operated agent workflows, simulated transaction states, or software configuration examples. AgentGuard does not transmit money, extend credit, custody funds, offer financial instruments, or guarantee settlement of obligations.

References to certify, verify, attest, signed, or validated are cryptographic and computational terms. They mean that software performed a signature, hash, schema, or chain check. AgentGuard does not claim that receipts are admissible legal evidence by default, that a court or regulator must accept them, or that use of AgentGuard alone fulfills any legal standard. The software is provided as-is as a technical audit tool.