Documentation answers. Human-reviewed moderation.
A self-hosted Discord bot that connects public documentation, moderation review and release announcements.
Documentation
Index configured public-document mounts with lexical retrieval. Hosted-document ingestion is an operator-controlled extension; document changes rebuild the index.
Human review
Borderline messages quarantine to a review card with Approve & Restore. Approved messages can be reposted through a per-channel webhook with a visible review reference. Recovery is a deliberate moderator action.
Three jobs, one agent.
Classification, grounded answers, and human review.Moderate
Per-message classifier with a six-rung action ladder (allow / warn / flag / delete / timeout / ban). Borderline messages soft-delete to a quarantine card; approved messages are reposted with a visible review reference. Timeout and ban verdicts are human recommendations.
Where the intelligence lives
Answer
Forum-aware support Q&A grounded on your docs. Retrieve top-k chunks → Sonnet with citations → reply in-thread. 👍/👎 reactions log feedback.
Announce
A webhook CLI (bizangent-announce) drops release notes into the
right channel from any of your repos. No bot-token sprawl. Wire it as a
Stop hook on tagged releases for operator-authorized release publishing.
Reference community: HeliosDB
Personal reference projectHeliosDB documentation and community workflows informed the project. This showcase describes the design; the examples below do not contact a live bot or model.
Illustrative documentation examples
No live model request; examples onlyA personal project, open to inspection.
Explore the code behind the examples.
Bizangent stays a personal, self-hosted project. There are no hosted plans, paid pilots or public customer registrations.
Data handling
Message text and retrieved public documents are sent to Anthropic for classification and answers. Provider processing and retention depend on the operator’s provider agreement; we do not promise zero retention. Restricted moderation records include text and supported attachments for recovery, ordinarily for 7 days; minimal audit metadata normally expires after 30 days. Remote review deletion is retried during outages; minimal references for unresolved delivery or cleanup remain until reconciled. Feedback stores answer/source identifiers and pseudonymous vote identity, not question bodies. Review cards and local records expire. This page only displays static examples; questions about a bot instance should go to its operator.