Every number, with its evidence.

Products in the dataset: the newest drive per product the directory shows.

Date of the first drive in the dataset

Date of the last drive in the dataset

Distinct categories among the products driven.

Drives an independent reviewer checked

Drives not yet approved for publication

Docs, sign-up, verification, key, first call, first action, payment, cancel, delete (/method).

L0 (Read-only): No API, CLI or MCP server. Agents can read about it, but there is nothing for them to use. (Source: /method.)

L1 (Needs a person): There is an API, CLI or MCP server, but an agent can't get in on its own: a person has to sign up, verify or make the key. What it does once a person sets it up shows next to the level. (Source: /method.)

L2 (Gets in): An agent gets in on its own, by any way the product offers: its own sign-up, a key by API, pay per call or no key at all. It does less than half the core work, or the work isn't measured yet. (Source: /method.)

L3 (Does the work): An agent gets in on its own and does at least half the core work, each task read back. (Source: /method.)

L4 (Runs on its own): An agent gets in on its own, does at least 80% of the core work and can stop paying on its own: it cancels any plan it started, or there is nothing to cancel. Whether it can also remove what it made shows next to the level. (Source: /method.)

L5 (Autonomous): Everything in L4, and the agent pays on its own for the work it does: a payment we made and checked on the product's agent payment route, such as x402 or MPP, that buys credits or a plan for its account, or pays per call for at least 80% of the core work. (Source: /method.)

L0 Read-only, L1 Needs a person, L2 Gets in, L3 Does the work, L4 Runs on its own, L5 Autonomous (Source: /method.)

Gets in on its own: sign-up, verify, credentials and first call pass through documented routes (or are n/a because a documented no-account route, e.g. a claimable database, gave working credentials). (Source: /method.)

Sign-up counts as open only when the newest drive got in and at least half of the last three did. Assisted drives are left out of this record. (Source: /method.)

Every product is entered one of three ways: our agent gets in on its own; a person made the account once (an assisted drive, or the owner's access on a paid drive), then the agent took over; or it does not get in, and the work inside is not measured. (Source: /method.)

Code computes levels and scores from the evidence; no person or agent types them in. (Source: /method.)

Our agent never solves a CAPTCHA, never works around a bot check or changes its behaviour to pass one, and never gives false information. When a check needs a person, someone from our team may clear that one step, and it counts as needing a human: the agent never gets credit for it. (Source: /method.)

No drive charges a card, unless the owner orders an Audit with their paid plan in it, at cost. The one payment a drive may make on its own is at most one dollar, through the product's own agent payment endpoint, and only if it has one. (Source: /method.)

Every claim in a report traces to a logged attempt. (Source: /method.)

Only documented routes count toward the level. (Source: /method.)

A drive has five parts: gets in and does the work, docs match, agent-ready setup, security and holds up. (Source: /method.)

Products where our agent signed up on its own (headless or in a browser)

Products where our agent signed up without a browser

Products where our agent signed up in a browser

Products whose sign-up needed a human

Products whose sign-up blocked our agent outright

Products where sign-up itself was the first checkpoint our agent did not pass on its own: it needed a human, or was blocked. Stops at the verification are not in it (the products stopped at the door at sign-up or its verification count the two together).

Products our agent got into on its own: sign-up, verification, a key and a first call with no person, on its newest drive and on at least half of its recent drives. Getting in in a browser counts when the API key needs a paid plan (the products it got into in a browser only).

Of the products our agent got into on its own, those where it got in in a browser while the API key needs a paid plan: in, but without an API key.

Every product our agent did not get into on its own (all products minus those it got into on its own): it needed a person once, or did not get in at all. Split by where its newest drive first stopped: at sign-up or its verification, at the key or the first call, or let in on its newest drive but not on enough of the ones before.

Products where a person made the account once and our agent went on from there. Part of the products stopped at the door.

Products our agent got into neither on its own nor with a person's help through documented routes. Part of the products stopped at the door.

Of the products stopped at the door, those whose newest drive first stopped at sign-up or verification (a step a person did counts as a stop).

Of the products stopped at the door, those whose newest drive got through sign-up and verification on its own and first stopped at the key or the first call.

Of the products stopped at the door, those whose newest drive got through the door on its own while fewer than half of its recent drives did.

Products where our agent got in on some recent drives and not on others

Products whose sign-up page we fetched as a signed agent Cloudflare already knows

Products whose sign-up page opened for a signed agent Cloudflare already knows, of those where we fetched it

Products that stopped agents on a server before the sign-up page loaded

Products that stopped agents on a server before the sign-up page and opened it for a signed agent Cloudflare already knows, of those that stopped them before the page

Products where robots.txt was what kept our agent out of sign-up

Products whose robots.txt closes the sign-up page to agents, whether or not it stopped ours

Products where docs was the first checkpoint our agent did not pass on its own

Products where sign-up itself was the first checkpoint our agent did not pass on its own: it needed a human, or was blocked. Stops at the verification are not in it (the products stopped at the door at sign-up or its verification count the two together).

Products where verification was the first checkpoint our agent did not pass on its own

Products where the key was the first checkpoint our agent did not pass on its own

Products where the first call was the first checkpoint our agent did not pass on its own

Products where the first action was the first checkpoint our agent did not pass on its own

Products where payment was the first checkpoint our agent did not pass on its own

Products where cancelling was the first checkpoint our agent did not pass on its own

Products where deleting the account was the first checkpoint our agent did not pass on its own

Products where our agent passed every checkpoint on its own

Products where our agent reached a verification step

Products where our agent verified the account on its own, of those with a verification step it reached

Products where our agent verified the account without a browser

Products whose verification stopped our agent

Products whose newest drive first stopped at sign-up, verification or the key (a step a person did counts as a stop): our agent had no API key of its own. Includes products it got into in a browser while the API key needs a paid plan.

Products whose newest drive passed sign-up, verification and the key on its own, or found them not needed (all products minus those stopped before a key).

Products where our agent got as far as the key

Products where our agent got a key on its own

Products where our agent got a key on its own, of those where it got that far

Products where our agent got a key without a browser

Keys our agent got without a browser, of all keys it got on its own

Products where the key had to be copied from a dashboard in a browser

Products where the key needed a browser and the first call then worked without one

Products where getting a key stopped our agent

Products where the key needed a human

Products where our agent could not get a key at all

Products where our agent got as far as a first API or MCP call

First calls that worked without a browser, of those our agent got to

Products where our agent's first API or MCP call worked

Products where our agent created something real on its own

Products where we measured how much of the core product our agent ran

Median share of the core product our agent ran, against the product's core features as mapped before the drive, over the products where it was measured.

Products where a bot check stopped our signed agent on a way in (sign-up to the first call), whether or not it got in another way: the ones it stayed out of plus the ones it got into on its own another way.

Of the products stopped at the door (our agent did not get in on its own), those where a bot check stopped our signed agent on its way in.

Products where a bot check stopped our signed agent on one way in and it got in on its own another way

Of the products where sign-up itself was the first checkpoint our agent did not pass on its own, those where a bot check at sign-up stopped our signed agent. Its total leaves out the stops at verification and the products with a bot check elsewhere.

Products whose bot check let our signed agent through

Products where our signed agent met a bot check on its way in (it stopped it or let it through)

Products where a bot check stopped our signed agent on at least one way in, of those where it met one

Bot checks that stopped our agent with a visible challenge

Bot checks that stopped our agent without showing any challenge

Bot checks that stopped our signed agent by judging its network and browser