What ships today
becos is a team- and tenant-scoped, server-backed record for AI coding agents. Everything below is live now — and every piece of it exists to make an agent's answer more accurate.
Grounded recall
vonic_query returns dated, attributed answers held to the record. Unsupported lines are dropped and the offered-vs-cited ratio is measured.
Resolvable citations
resolve_event returns a cited event's stored source verbatim — no summarisation, no model call.
Currency & contradiction
Superseded values come back marked superseded; a draft that conflicts with a recorded decision is surfaced against it.
Structural tenant isolation
Results are filtered by tenant prefix before any model sees them. Cross-tenant access is closed by construction, not policy.
Injection-safe by construction
The record is treated as untrusted input. Its content can't express slug / path / tool fields, and nothing that writes the record holds a tool.
Entity knowledge graph
Repository, branch, feature and author pages with typed links and rename-safe identity keyed on stable external IDs. A feature_registry prevents near-duplicate pages.
One record across Claude Code, Cursor, ChatGPT and opencode — via the becos connector, not a single vendor's plugin.
Built as a record, not a cache
changed_files[] with add/del/status, repositories_touched[]. No diffs, no tool_use, no thinking.event_id. Lifecycle labels mean an answer never presents a pending or failed entry as settled truth. Failures are stored content-free — exception type, never the message.What's roadmap, said plainly
Decision archaeology
Reconstruct the state of a decision at any past date — “what would this answer have been then.” Designed, substrate present, not shipped.
Layered recall
Broad cheap triage first, full cited evidence only for the facts that survive — keeps grounding fast over long histories. In design.
See what it recalls on your repo
Connect one repository and ask your agents a question you'd normally interrupt someone for.
Get started