Every external capability is a contract.
Supabase, Follow Up Boss, Notion, Drive, semantic document retrieval, reviewed Slack memory, skills, saved queries, and market recommendations are exposed through 27 named tools.
Synthetic product demonstration
The operating agent, inside Slack.
Choose a scenario to see the question, the bounded tools selected, the evidence returned, and the answer delivered in-thread. Names, records, and values are synthetic; the tool contracts and control model match the implemented system.
27Tool contracts
8Maximum calls
87/87Tests passing
1Audit record per run
# operations
Synthetic data
What the simulation represents
Useful autonomy with explicit boundaries.
Eight calls means eight calls.
The model can investigate across sources, but the runtime owns the call budget, timeout, retry behavior, within-run cache, result limits, and final audit write.
Read-only survives a bad prompt.
A SELECT guard, read-only Postgres role, transaction settings, statement timeout, approved schema, and result caps enforce the boundary below model instructions.
Learned state stays inspectable.
Skills have names, descriptions, categories, and workspace scope. Slack guidance is review-gated. Short or zero-tool exchanges do not trigger low-value memory writes.
A question can become an operation.
Saved queries store the original question, target channel, validated cron, and optional read-only condition. The scheduler can skip model cost and Slack delivery when nothing changed.
The claim is tied to the current code.
The current registry contains 27 tools. The focused agent suite passed 87 of 87 tests on July 19, 2026. This page exposes no credentials, client records, or raw private messages.