Local-First Data Broker Removal: Extraction, Deletion, and Rechecks

What it is

unbroker is an optional Hermes Agent security skill. Its README describes a deterministic Python CLI that owns the state: config, dossiers, broker data, tier planning, ledger, drafts, reports, and the action queue.

The agent side does the browser and extraction work. That split is the interesting part: the model does not improvise the entire privacy workflow from scratch. It follows a queue, records outcomes, and escalates hard cases into one digest.

What to check

A removal agent handles the same personal data it is trying to suppress. Check consent, local storage, disclosure rules, verification, human fallback, and re-scan timing before using one.

Keep broker removal separate from payment automation. This is a personal-data workflow. Its main risk is unnecessary collection or disclosure, not payment metadata.

• Get explicit consent before a scan or removal request.
• Keep the person record and ledger local where possible.
• Send only the fields required by the removal channel.
• Stop for CAPTCHA, ID, phone, or other human-only steps.
• Check the official repository for changes to broker coverage and workflow rules.

What matters operationally

Keep autonomy constrained. unbroker says consent is required, hard CAPTCHA and government-ID cases become human tasks, and confirmed removal should only be recorded after a verifying re-scan.

Data-broker cleanup is repetitive, perishable, and full of hostile forms. A local ledger, narrow disclosure, and recurring re-scan loop are more useful than a one-shot deletion claim.

How DROP changes removal work in California

California DROP is now a separate path for California residents. The official state page says a California resident can submit one request to more than 600 registered brokers, and that brokers must begin processing DROP deletion requests on August 1, 2026.

That means the best workflow for a California resident is not only per-site scraping. It is DROP first, then people-search parent clusters, then standalone opt-outs, then scheduled re-checks.

Separate extraction from removal

Extraction should produce evidence, not a deletion claim: source URL, matched subject fields, capture time, and the broker path that appears to control the record. Keep this local and verify that the record belongs to the consenting subject before preparing a request.

Removal is an external disclosure. Draft only the fields required by the broker, stop for review before sending, and keep government ID, phone numbers, and unrelated aliases out of the payload unless the specific verified route requires them. A submitted form remains pending until email confirmation and a later lookup support the result.

• Extract and normalize evidence locally.
• Match the record to the authorized subject.
• Draft the minimum required removal payload.
• Require approval before browser or email submission.
• Store confirmation IDs, timestamps, and status without retaining unnecessary page data.
• Recheck on a schedule because brokers can repopulate records.

Self-hosting and fallback boundaries

A local machine is the clean default for personal runs because dossiers, the task ledger, browser sessions, and optional local-model prompts stay close to the operator. Bind interfaces to localhost or a private network, keep secrets outside the repository, use a separate workspace per subject, and encrypt the host storage.

A temporary VPS can isolate a run, but it creates a new custody boundary for personal data, logs, backups, remote access, and deletion. Whichever host is used, submission still contacts broker sites, mail providers, CAPTCHA systems, or an optional remote model. Self-hosting does not erase those outbound disclosures.

Adapters should be an allowlist of expected domains, form fields, actions, and evidence. If a broker changes its flow, fail closed and return the job to the human digest rather than allowing a general browser agent to guess through it.

Limits

Broker removal is not erasure. Public records, court files, property records, paid-tier retention, and re-listing can remain. A removal agent reduces exposure; it does not delete reality.

The workflow is built around subject authorization and minimized disclosure. The same technical shape can also support people-search profiling, so the intake and ledger boundaries matter.