Browse documentation
Documentation/Agent classification

Agent classification

Read experimental human, assistant and script scores without confusing them with verified identities.

Doorman separates who your app authenticated from what the current activity resembles. A verified agent credential is an identity fact. A model’s assistant score is an estimate. Neither is a verdict that the activity is abusive.

Verified people and agents

Pass the actor your server authenticated through the normal endpoint:

const result = await doorman.assess(request, {
  auth: {
    userId: owner.id,
    accountId: workspace.id,
    actor: { id: authenticatedAgent.id, kind: "agent" },
  },
});
return result.response;

Your application must verify that credential and its permission to act for the owner. For a person acting as themselves, omit actor; the authenticated user is the actor. Count these verified IDs in account reports, keeping browsers separate from people.

Optional activity estimates

Enable Jev on your server and, if your collection policy permits, add extended browser summaries:

const doorman = createDoormanClient({
  endpoint: "/api/visitor",
  collection: "extended",
});

The recommended server API requests human, assistant and scripted-automation scores when enough aggregate evidence is present. No separate operator service, training dataset or classifier deployment is needed.

Read result.identity?.operator privately on the server:

Field Meaning
status evaluated, insufficient-evidence, unavailable or disabled.
label human, assistant, automation or unknown.
scores Available human, assistant and automation estimates from 0 to 1. They need not sum to one.
calibrated false: these are not measured real-world probabilities.

The label can remain unknown even after evaluation. Sparse input, weak scores and close alternatives should not force a label. Set classifyOperator: false on the TypeScript server to disable this optional classification.

The separate risk.automation and risk.suspicious scores describe technical automation and suspicious activity. A legitimate assistant can score highly on automation without being suspicious. Server request aggregates can add evidence of denied operations without claiming a person is an attacker.

What you can report

Use result.properties with your existing server analytics events. Report evaluated activity by label, and show unknown/unavailable results separately. Keep that distinct from the number of authenticated agent credentials or users.

Doorman cannot reliably tell you that one shared password belongs to a husband and wife, count physical people, or identify ChatGPT versus Claude from mouse movement. Our recorded 80-session Jev study did not establish reliable human/agent separation. Results and limitations.

Missing mouse movement, privacy settings, software rendering, aligned clicks or a focus change are not proof of automation. Doorman respects hidden browser values and stores aggregate summaries rather than keys, forms or coordinate trails. See optional signals.

Search documentation

Search setup instructions, examples and the API reference.

Local search. No query leaves your browser.