What are the main AI agent detection tools?
AI agent detection tools tell you which automated client sent a request, and the main ones as of November 2026 are Cloudflare, Akamai, DataDome, HUMAN, Fastly, Vercel BotID, and Arcjet. Most run at the network edge or in a module in front of your application, while BotID and Arcjet return their result to your request handler. They differ on three questions: whether they separate an AI agent acting for a waiting person from an AI crawler collecting pages ahead of time, how they verify a bot's claimed identity, and whether they tell your code how strong that verification was.
All seven separate AI agents from AI crawlers in some form, and six of the seven, including Arcjet, document verifying Web Bot Auth signatures. Only Arcjet documents verifying the Personal Agent Protocol (Poppy), which shows that an agent acts for one of its users. Edge products fit volumetric attacks and a whole-site default with no code change. Arcjet fits per-route decisions that need application context, such as the signed-in user, the cart, or the plan. Arcjet doesn't replace network-layer DDoS protection.
Arcjet publishes this comparison. Competitor details come from each vendor's public documentation and official blog, and from OpenAI's published agent access guide, reviewed on November 16, 2026. Where a vendor's documentation doesn't state a capability, the table says "Not documented", which doesn't mean the product lacks it. Products change, so check the linked sources before you decide.
What should an AI agent detection tool tell you?
An AI agent detection tool needs to answer three questions: which bot sent the request, how sure the tool is, and whether the bot acts for one of your users. The old bot question – person or machine – doesn't help when a customer sends ChatGPT agent or Meta's Muse to book a table, because a machine sends the request for a paying customer.
The following terms describe what each tool can report:
- AI agent versus AI crawler. An AI agent fetches a page on demand for a person who is waiting on the answer. An AI crawler collects pages ahead of time, for model training or a search index. The same operator can run both, such as OpenAI's
ChatGPT-UserandGPTBot. For more information about the split, see how to block AI crawlers but allow AI agents. - Verification method. A
User-Agentheader costs nothing to forge. IP ranges and forward-confirmed reverse DNS show that a request left the operator's network. A Web Bot Auth signature, built on RFC 9421 HTTP Message Signatures, shows that the sender controls a key the operator published. - Delegation. A Poppy session shows that a vetted agent acts for one of its users, without identifying the user.
- Evidence level. A verdict such as "verified bot" tells you the outcome. An evidence level tells you which of those methods produced it, so you can require a signature on checkout and accept an IP match on a product page. For more information, see AI agent evidence levels.
How do AI agent detection tools compare?
The tools differ most on Poppy and on how much of the verification result reaches your code. The following table compares how each tool identifies AI agents, as of November 2026.
| Tool | Separates AI agents from AI crawlers | Verifies Web Bot Auth | Verifies Personal Agent Protocol | Reports an evidence or verification level |
|---|---|---|---|---|
| Cloudflare | Yes. Search, Agent, and Training behaviors in AI Crawler and Agent access policies , and AI Assistant and AI Crawler verified bot categories . | Yes, in Verified Bots and signed agents. | Not documented. |
Bot Management fields . A verification method field isn't documented. |
| Akamai | Yes. Three AI categories : training crawlers, search crawlers, and fetchers and agents. | Yes, per Akamai and OpenAI . | Not documented. | A bot score for each request . A verification level isn't documented. |
| DataDome | Yes. AI Crawler, AI Assistant, Autonomous Agent, and Agentic Browser . | Yes, per its bot authentication docs . | Not documented. | High, medium, or low identification strength for each agent, shown in dashboards. A request header for it isn't documented. |
| HUMAN | Yes. AI agents have their own permissions , and AI crawlers are managed in traffic policy settings. | Yes, per HUMAN and OpenAI . | Not documented. | High, medium, or low trust level for each agent. Per-request exposure isn't documented. |
| Fastly | Yes. AI crawler and AI fetcher signals . | Not documented. | Not documented. | Verified or suspected, as separate signals . |
| Vercel BotID | Yes. Directory categories include | Yes, per Vercel and OpenAI . | Not documented. | An
. A verification method field isn't documented. |
| Arcjet | Yes. bot identity . | Yes, for catalog signers and any directory you name . | Yes, for agents vetted in the catalog . |
|
The tools also differ on where they run, whether their bot list is public, and how they charge. The following table compares those three areas.
| Tool | Where it runs | Open catalog | Pricing model |
|---|---|---|---|
| Cloudflare | Cloudflare's edge. Traffic must go through Cloudflare. | Public bots and agents directory on Cloudflare Radar. Open source isn't documented. | AI Crawl Control on all plans . Bot Management is an Enterprise subscription. |
| Akamai | Not documented as public. | No published price. Demo and contact sales . | |
| DataDome | CDN, web server, or backend module , with edge recommended. | Not documented as public. | Plans linked to a number of requests . |
| HUMAN | An enforcer on a CDN, load balancer, or origin , plus a browser sensor. | Not documented as public. | Not documented. |
| Fastly | Fastly's edge . Bot signals need an Edge WAF deployment. | Not documented as public. | Monthly request volume, per million . |
| Vercel BotID | A browser challenge plus | Public directory at bots.fyi. Open source isn't documented. | Basic free; Deep Analysis $1 per 1,000 calls on Pro . |
| Arcjet | Your request handler on any host, with SDKs for JavaScript, Python, and Go. | Open source well-known-bots catalog. | Plan plus $5 per million web requests. See Arcjet pricing. |
How does Cloudflare detect AI agents?
Cloudflare identifies AI agents through Verified Bots, which accepts a Web Bot Auth signature, a published IP list with a stable User-Agent, or reverse DNS as proof of identity. Cloudflare announced signed agents as a classification for user-directed agents that sign with Web Bot Auth. Its Verified Bots documentation states that, as of July 1, 2026, the difference between a verified bot and a signed agent is a Direct or Intermediary access field: an intermediary is an agentic service that many end users operate.
Cloudflare separates AI traffic by behavior. AI Crawler and Agent access policies apply Search, Agent, and Training actions, where Agent covers "automated activity acting in real time on a person's behalf". The policies apply to verified bots and to matching unverified bots, and Cloudflare's docs describe them as available to all customers. AI Crawl Control, on all plans, adds per-crawler allow and block rules and robots.txt compliance tracking.
In rules and Workers, Bot Management fields include cf.bot_management.verified_bot, cf.bot_management.signed_agent, cf.verified_bot_category, and a bot score from 1 to 99. The docs don't list a field that names the verification method. Cloudflare is listed as a company design partner on the Personal Agent Protocol's path to 1.0, but its bot documentation doesn't document verifying Poppy sessions, as of November 2026.
Cloudflare's strengths are scale and default coverage: unmetered DDoS mitigation and zone-wide AI crawler policy with no code change. For more information, see Cloudflare vs Arcjet.
How does Akamai detect AI agents?
Akamai verifies Web Bot Auth signatures at its edge and classifies AI traffic in its bot directory. Akamai's November 2025 post states that "Web Bot Auth is already supported by Akamai's bot & abuse solutions" and that its edge validates signatures using RFC 9421 rules (Akamai). OpenAI's agent access guide lists Akamai among the providers that verify ChatGPT agent's signatures automatically.
Akamai's October 2026 post says Akamai split its AI bot category into AI training crawlers, AI search crawlers, and AI fetchers and agents, and that "any operator can register with Akamai and have its traffic verified" (Akamai). The Bot Manager product page describes edge mitigation, a bot score for each request, and a known bot directory that you can extend with your own categories. Poppy verification and a per-request verification level aren't documented. Akamai doesn't publish a price for Bot Manager.
How does DataDome detect AI agents?
DataDome identifies AI agents with methods it rates by strength and puts them in four categories: AI Crawler, AI Assistant, Autonomous Agent, and Agentic Browser (DataDome). High-strength methods are Web Bot Auth, cryptographic token verification, trusted IP lists, and reverse DNS. Network attribution and client-side fingerprinting are medium, and server-side fingerprinting is low. DataDome shows identification strength in its Access Control and Explore dashboards.
DataDome's bot authentication docs state that it "can cryptographically verify http signatures using the Web Bot Auth protocol", and that unauthenticated automated requests are categorized as Threat Detection and blocked by default. Access Control sets Intent-Based, Allow, Block, Rate Limit, Time-Box, or Monetize policies by agent, endpoint, or custom rule. Poppy verification isn't documented. For a full comparison, see DataDome vs Arcjet.
How does HUMAN detect AI agents?
HUMAN assigns each AI agent a high, medium, or low trust level, and gives a high level when it can "verify the agent via cryptographic key on each request" (HUMAN). The same page points to Web Bot Auth for cryptographic authentication, and OpenAI's agent access guide lists HUMAN among the providers that verify ChatGPT agent's signatures automatically. By default, an agent starts at medium.
HUMAN's agent permissions go further than allow or deny. Each agent has separate settings for reading content and products, engaging, account creation, login, account changes, checkout, and a per-minute rate limit. AI crawlers are managed separately, in traffic policy settings. HUMAN's enforcers run on a CDN, load balancer, or origin, with frameworks from Cloudflare Workers and Fastly to Node.js, Python, and Go. Poppy verification and pricing aren't documented.
How does Fastly detect AI agents?
Fastly detects AI bots at its edge with four signals that separate crawlers from fetchers and verified bots from suspected ones. VERIFIED-BOT.AI-CRAWLER covers crawlers "used for training AIs and LLMs", and VERIFIED-BOT.AI-FETCHER covers fetchers "for enriching results in response to a user query" (Fastly). The SUSPECTED-BOT variants cover the same two groups without verification. Bot signals are available to Edge WAF deployments, and the AI Bot Management page states that "all bot detection and mitigation occurs at the edge".
Fastly's documentation doesn't describe how it verifies bots, and Web Bot Auth verification isn't documented. In August 2025, Fastly wrote that it was exploring the proposal (Fastly). Fastly charges for Bot Management by monthly request volume.
How does Vercel BotID detect AI agents?
Vercel BotID identifies AI agents through Vercel's verified bot directory: checkBotId() returns isVerifiedBot, verifiedBotName, and verifiedBotCategory, and your code branches on them (Vercel). Vercel's bot management docs list IP address verification, reverse DNS, and Web Bot Auth as verification methods, and its directory categories include ai_assistant (such as chatgpt-operator), agent (such as google-agent), and ai_crawler (such as gptbot).
Vercel announced Web Bot Auth support in August 2025. The verified-bot fields report whether a bot is verified, not which method verified it, and Poppy verification isn't documented. BotID itself is a browser challenge for projects deployed on Vercel. For more information, see Vercel BotID vs Arcjet.
How does Arcjet detect AI agents?
Arcjet reports one primary bot identity on each bot decision, with a kind, a bot ID, and the evidence that established it. The kind is AI_AGENT for a bot that acts on demand for a waiting person, AI_CRAWLER for one that collects pages ahead of time, AUTOMATION for browser automation that advanced signals detected with no named bot, or OTHER. The evidence runs from weakest to strongest: UNVERIFIED (a User-Agent claim or IP reputation), FINGERPRINT (advanced signals), IP (published ranges or forward-confirmed reverse DNS), and SIGNATURE.
SIGNATURE evidence comes from a verified Web Bot Auth signature or a verified Poppy session. Arcjet verifies the signers in the open well-known-bots catalog, such as CHATGPT_AGENT, GOOGLE_AGENT, BROWSERBASE, and META_MUSE, and any other signer you name with SIGNED_BOT: and its directory URL. A verified Poppy session sets delegated to true and operator to the agent's domain. Arcjet verifies sessions only for agents vetted in the catalog, so a request can't make Arcjet fetch a client_id URL it supplies. When IP or reverse DNS checks contradict a claim, spoofed is true.
The decision runs in your handler, so the identity sits next to your own data. You can allow CATEGORY:AI_AGENT on product pages, deny CATEGORY:AI crawlers, require SIGNATURE on checkout, and give a delegated agent with a valid OAuth token access to the signed-in user's cart. For the API, see bot identity in the Arcjet docs.
Arcjet's limits are plain. Only Arcjet Cloud computes the identity, so a decision the SDK makes locally has none. A signature proves the operator of a key directory, not the end user: Browserbase and Kernel sign for many customers. A Poppy session doesn't prove that the user approved a specific purchase. Arcjet runs inside your application, so it doesn't absorb a network-layer DDoS attack.
When does an edge product fit better than in-code detection?
An edge product fits better when the threat is volume, when you want one default for every route, or when you can't change application code. Cloudflare, Akamai, and Fastly absorb traffic before it reaches your origin, which in-code detection can't do. Their AI crawler policies apply to a whole zone in one setting, and they protect applications that you can't modify, such as a vendor-hosted storefront.
In-code detection fits when the right answer depends on the request's context. An edge rule can admit ChatGPT agent, but it doesn't know whether the agent is about to change the shipping address on an account with a large unshipped order. Arcjet returns the identity in the same handler that holds the user, the cart, and the plan, so one function makes both decisions. For more information about why identity and authorization stay separate, see agentic commerce bot protection.
Many teams run both. A CDN handles DDoS and coarse crawler policy, and Arcjet runs behind it for per-route bot rules, evidence checks, and rate limits keyed on the user.
What can't any AI agent detection tool see?
No server-side tool can identify an AI agent that drives a person's own signed-in browser, such as Claude in Chrome or Perplexity Comet. Those requests carry the person's cookies, IP address, and browser fingerprint, and send no signature. Detection tools can still catch automation in the page through client signals, and rate limits slow actions that people don't repeat at machine speed. For more information, see how to detect AI browser agents.
Verification also has a ceiling. A Web Bot Auth signature identifies an operator, a Poppy session identifies an agent and shows that it acts for a user, and neither proves intent. Authorization for purchases and account changes stays in your application.
Which AI agent detection tool fits your job?
Choose by where the decision belongs and how much of the verification result your code needs. The following sections list the signals for each choice.
Choose an edge bot manager
Choose Cloudflare, Akamai, or Fastly if any of the following apply:
- You need network-layer DDoS mitigation with bot management in the same product.
- You want one AI crawler policy across every route with no code change.
- You protect applications whose code you can't change.
- Your traffic already routes through that CDN.
Choose DataDome or HUMAN
Choose DataDome or HUMAN if any of the following apply:
- You want a fraud and bot team's dashboard with per-agent policies set outside the code.
- You want integrations across many CDNs, web servers, and mobile SDKs.
- You want per-agent permissions such as checkout or login managed by a security team.
Choose Vercel BotID
Choose Vercel BotID if you deploy a JavaScript application on Vercel and want a browser challenge plus verified-bot fields with minimal setup.
Choose Arcjet
Choose Arcjet if any of the following apply:
- You need per-route decisions that use the user, the cart, or the plan.
- You want the evidence level and the
delegatedflag in your code, not only a verified-bot verdict. - You want Web Bot Auth and Poppy verified and reported in one identity, so you write one policy instead of two integrations.
- You want an open catalog that you can read and contribute to.
- You use Python or Go as well as JavaScript, on any host.
For the standards behind the identity, see Web Bot Auth vs the Personal Agent Protocol. For how Arcjet identifies specific agents, see how to identify AI agents and bots.
Frequently asked questions
What is an AI agent detection tool?
An AI agent detection tool identifies which automated client sent a request and whether it's an AI agent acting on demand for a person or an AI crawler collecting pages for training or a search index. The useful ones also verify the claim, with IP ranges, reverse DNS, or a Web Bot Auth signature, rather than trusting the User-Agent header.
Which bot management products verify Web Bot Auth?
As of November 2026, Cloudflare, Akamai, DataDome, HUMAN, Vercel, and Arcjet document verifying Web Bot Auth signatures, and OpenAI's agent access guide lists Akamai, Cloudflare, HUMAN, and Vercel as verifying ChatGPT agent's signatures automatically. Fastly's documentation doesn't describe Web Bot Auth verification; in August 2025 Fastly wrote that it was exploring the proposal.
Which tools verify the Personal Agent Protocol?
As of November 2026, Arcjet documents verifying Personal Agent Protocol (Poppy) sessions for agents vetted in its open well-known-bots catalog, and reports them with SIGNATURE evidence and delegated set to true. Cloudflare is listed as a Poppy design partner, but its bot documentation doesn't describe verifying sessions. The other tools' documentation doesn't mention Poppy.
What's the difference between a verified bot and a signed agent?
A verified bot is one whose identity a provider has confirmed, by IP ranges, reverse DNS, or a cryptographic signature. A signed agent is an agent verified by a Web Bot Auth signature. Cloudflare's docs state that, as of July 1, 2026, signed agents are verified bots with a Direct or Intermediary access field. Arcjet reports the method as the evidence level: IP or SIGNATURE.
Should I use an edge bot manager or in-code detection?
Use an edge bot manager such as Cloudflare, Akamai, or Fastly for volumetric attacks, a whole-site default, and applications you can't change. Use in-code detection such as Arcjet when the decision depends on the signed-in user, the cart, or the plan. Many teams run both, with Arcjet behind the CDN. Arcjet doesn't replace network-layer DDoS protection.
Can any tool detect AI agents that run in a person's browser?
No server-side tool can identify an agent such as Claude in Chrome or Perplexity Comet that drives a person's own signed-in browser, because the requests carry that person's cookies, IP address, and fingerprint. Client-side automation signals and rate limits still help on forms where automation matters.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Arcjet runs inside your application, where it can use runtime context to enforce agent actions and budgets, detect prompt injection, and protect sensitive information before a workflow acts.