Most AI tools take a question and hand back an answer. You have no idea how it got there, what it relied on, or whether it made anything up. IKKF shows its working.
Every run moves through Intent, Retrieval, Decomposition, Reasoning, Verification, Composition and Confidence, streaming each phase as it happens. You watch the reasoning build, then inspect the result.
Cross-checking runs against your own knowledge base. Each atomic fact is a Markdown file with a claim and a source, stored on your Mac and never uploaded. The verification phase checks claims against those sources, and flags what it cannot support.
Most confidence scores are the model's opinion of itself, and change run to run. IKKF computes its score from countable evidence: how many facts were retrieved, how many reasoning steps cite a source, whether verification passed, and how many uncited claims remain. The same inputs give the same score, so runs are comparable. The model's own judgement is shown alongside, clearly labelled as an opinion.
Bring any OpenAI-compatible endpoint — Ollama or LM Studio on your own machine, or a cloud provider. Local models need no account and no key.
When a run completes, IKKF distils the experience into a reusable skill, saved to your library and written to disk as a standard SKILL.md. Re-running a task versions the skill rather than duplicating it.
IKKF is not a chatbot and not a cloud service. It is a reasoning workbench for people who want to see how an answer was reached. It requires a model to think with — a local one, or an endpoint you provide. It runs on macOS.