CallChat ZERO: Consent-First AI Lead Agent

A privacy-preserving AI chat and lead agent with explicit consent, protected PII, abuse controls, automatic retention limits, and auditable human handoff.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CallChat ZERO: Consent-First AI Lead Agent

Inspiration

Private communication and AI-assisted lead capture should not require people to surrender control of their identity. CallChat ZERO starts from a simple rule: consent, data minimisation, and human control are product features—not policy footnotes.

What it does

CallChat combines a self-hosted Matrix communication stack with a consent-first AI lead agent. Visitors can use encrypted chat and calling, while the lead flow collects only the information a person explicitly chooses to provide: a required reply email plus optional name, organisation, and coarse city/country. It avoids precise location and IP-based profiling, adds abuse controls, supports auditable human handoff, and applies an automatic retention window.

How we built it

The communication layer uses Matrix, Synapse, PostgreSQL, Element Web, MatrixRTC, and Matrix end-to-end encryption. A Python agent bridge handles the lead workflow. During Build Week, Codex and GPT-5.6 were used to inspect the existing architecture and implement a focused privacy and reliability patch: consent-aware collection, honeypot and rate-limit protections, minimised notification content, 30-day cleanup, and accessibility-safe modal behaviour. Six targeted tests pass locally.

The wider CallChat research surface includes ZShield/ZME1 experiments. These are labelled as research layers and are not presented as replacements for established cryptography.

Challenges

The hardest part was keeping the agent genuinely useful while collecting less data. We also had to prevent anti-abuse controls from harming legitimate visitors, preserve keyboard and screen-reader behaviour, and keep experimental security work clearly separated from verified protections.

Accomplishments

  • Consent-first lead capture with an intentionally small data footprint
  • Self-hosted, owner-controlled communications infrastructure
  • Abuse controls and retention limits built into the workflow
  • Clear separation between Matrix E2EE and experimental security research
  • A tested path from AI conversation to accountable human follow-up

What we learned

Privacy is not one checkbox. It is a chain of decisions covering collection, storage, notifications, retention, deletion, accessibility, and operator behaviour. Codex and GPT-5.6 helped turn that principle into a reviewable set of concrete changes.

What's next

Next steps are a public Build Week README, a short evaluator-focused demo, broader integration testing, and independent review of the research security layers.

Built With

Share this project:

Updates