kitsokiPut the workflow back in charge of the agent
Kitsoki runs conversational workflows as auditable state machines. The LLM is a bounded callee at named decision points, not the hidden dispatcher for every turn.
Kitsoki runs conversational workflows as auditable state machines. The LLM is a bounded callee at named decision points, not the hidden dispatcher for every turn.
onboard ., review the profile, and commit the setup Kitsoki writes.How do I author and test stories?Use the docs by task: story authoring, replay without live LLM spend, debugging, and architecture.Kitsoki is for workflows that repeat, need human gates, or have to be defended after the run. The runtime owns the state machine. The model is called only at declared decision points and returns a typed result. Deterministic turns route without a model call; ambiguous turns, host calls, guardrail retries, and world mutations land in a structured trace.
The same app.yaml drives the web UI, terminal UI, MCP studio surface, headless runs, docs-driven demos, and replay fixtures. The product surface is not a mockup of the workflow; it is the workflow.
These are the demos a curious developer should watch first. The full feature and story catalog is still available from Demos and features.

Every agent call's full tool-use stream — captured once, replayed byte-identically.
Verdicts, confidence, evidence, and replay — on a frozen trace.

Agent questions become blocking operator modals — never silent defaults.

Triage a ticket, hand it to the pipeline, watch it land.
A released binary, dropped into a foreign repo, profiles it and installs the toolkit without a checkout.