kind and a value, and this page lists the values each kind can take. For where attribution appears in the API and webhooks, and how its shape differs between them, see Verdicts & scoring. The same lists are available as JSON at attribution.json.
Foil adds values as it learns to recognize new clients, and adding a value is not a breaking change. Treat
value as an open string: store values you don’t recognize, and include a default branch wherever you switch on one. See Versioning.provider: meta-muse for the hosted browser and actor: ai-agent for the type of client. To find a particular AI agent, match on the labels in AI agents and agent browsers.
Actor
Theactor label says what type of client is behind the session.
A verified Web Bot Auth signature doesn’t set an
actor value. It adds trust: verified and names the signing domain in a product label, and the actor comes from what the signer is, such as ai-agent for an AI agent’s signing domain. To find verified bots, match on trust: verified.
Provider
Theprovider label names the hosted browser, scraping service, or device farm the session ran in, when Foil recognizes it.
Foil recognizes a provider from the browser it runs, and for some providers also from the network the session leaves through. The same browser on another network then carries no provider label. Firebase Test Lab is named from its network alone. A provider label doesn’t set the verdict: a session from a recognized provider can still score inconclusive, so decide on the verdict and treat the label as context.
Sessions scored before October 29, 2026 can carry
chatgpt-agent for Dot’s hosted browser.
AI agents and agent browsers
Foil names an AI agent or agent browser with aprovider or a product label. This table lists the labels that name each one and the evidence behind them. A signature is a verified Web Bot Auth signature on the request.
A hosted browser that runs only one AI agent also adds actor: ai-agent. These are the hosted browsers of Muse, Instinct, Dot, Grok Bot, Grok, Perplexity Computer, and Hark. A hosted browser that also runs scrapers, such as Browserbase or Kernel, names only the platform and doesn’t add an actor label. Pally runs in Kernel’s browser with a setup of its own, so a session with that setup is named pally and also carries actor: ai-agent.
product: comet names the browser, not who was using it, so a person browsing in Comet carries it too. It comes with organization: Perplexity and no actor label. Foil can name Comet only where its script runs in the top-level page, not inside an embedded frame.
product: aside identifies Aside from its bundled browser runtime. It carries no actor label because a person browsing in Aside produces the same evidence as its assistant. Missing or blocked runtime evidence is inconclusive.
The AI coding app labels, from product: cursor-browser to product: tenslo-code-browser, work the same way. Each names the app’s built-in browser or web preview from the user agent it sends, which the app’s agent and the developer both use, so none comes with an actor label. Where a company makes the app, the label comes with its organization, such as organization: Anysphere for Cursor or organization: Microsoft for Visual Studio Code. The user agent reaches embedded frames, so Foil names these browsers there too.
Crawlers that send a browser user agent
Some crawlers, such as the ones that build link previews and review ads, render pages in a full browser and send that browser’s ordinary user agent, so nothing in the request names them. Foil names them from the network the browser connects from when that network belongs to the crawler’s operator.
Foil adds these labels only when no other label already names the client. A session whose user agent names the crawler, such as Googlebot or facebookexternalhit, keeps the labels from its user agent instead, and a session with a
provider label or an AI agent label doesn’t get them. Google’s ranges leave out Google Cloud customers, the fetchers Google runs at a user’s request, and Google-Agent, so a site hosted on Google Cloud or a person on a Google-run VPN or relay doesn’t get these labels. Muse connects through other networks and doesn’t get them either.
These labels don’t change the verdict. They can appear next to actor: automation and a tool label, which describe how the crawler drove its browser.
Monitors and test devices named by their network
Some synthetic monitors run an ordinary browser and send its usual user agent, and apps under test run on phones in a device farm. Foil names both from the network they connect from.
Firebase Test Lab also runs Google Play’s pre-launch reports, which test every app uploaded to a testing track. A login, signup, or Plaid Link flow inside your app can then reach Foil from Google’s test devices. Test Lab runs scripted tests and automated crawls, and developers can also use its phones by hand, so these labels carry no
actor.
Like the crawler labels, these labels don’t change the verdict and are added only when no other label already names the client. A monitor whose user agent names it, such as Datadog’s DatadogSynthetics, Dynatrace’s RuxitSynthetic (product: dynatrace-synthetics), or WebPageTest’s PTST (product: webpagetest), is named from its user agent instead.
Apps and browsers
Many apps open links in a built-in browser, and some browsers add their own name to the user agent. Foil names the app or browser from that user agent with aproduct label and the maker’s organization label. An app’s built-in browser also carries environment: in-app-browser.
These labels name where the page ran, not who ran it, so they carry no
actor. They describe a person’s browser, so Foil leaves them off a session scored bot and off one that another label already names as an automated client, AI agent, crawler, or hosted browser. chatgpt-desktop-app is the ChatGPT app’s own window, such as a bank-linking flow ChatGPT opens. It isn’t ChatGPT’s agent, whose browser is chatgpt-in-app-browser.
Tool
Thetool label names the automation framework or client library.
Product
Theproduct label names the specific bot, agent, or product. Its value comes from one of these sources:
- The product token from the client’s user agent, lowercased with hyphens, such as
googlebot,bingbot, ormanus-user. - For many well-known crawlers, a family name and a product name, sent as two separate
productlabels.GoogleOtherarrives asproduct: googleandproduct: other,GPTBotasproduct: openaiandproduct: gptbot, andOAI-SearchBotasproduct: openaiandproduct: searchbot. The second name can repeat across families, assearchbotdoes for OpenAI and Anthropic, so match on both labels. - For a request with a verified Web Bot Auth signature, the signing domain, such as
chatgpt.com. - For computer use, the browser the agent drove, such as
claude-in-chrome. - For an AI browser or an AI app’s built-in browser, the browser itself, such as
cometorcursor-browser. The value names the browser even when it comes from a token in the user agent. - For another app’s built-in browser or a browser that names itself in its user agent, the app or browser, such as
facebook-apporduckduckgo-browser. See Apps and browsers.
product value. Foil recognizes more than 1,500 products, so treat product as an open string. AI agents and agent browsers lists the product values that name AI agents.
Organization
Theorganization label names the company that operates the client. Its value is the company’s display name, such as OpenAI, Google, or Anthropic, rather than a lowercase identifier. Treat organization as an open string.
Purpose
Thepurpose label says why a crawler or fetcher is on your site.
Environment
Theenvironment label describes the browser engine and where the browser runs.
Trust
Thetrust label says whether the client’s claimed identity holds up.
Evasion
Theevasion label says the client tried to hide that it is automated.