Kairos — Private AI Brain Assistant
A private desktop AI system that runs models locally through Ollama, manages external provider APIs deliberately, and routes every answer through bounded, cited context.
Runtime and tools
- Ollama
- Qwen 3.6
- Tauri
- Rust
- Codex CLI
- Claude Code
Product at a glance
- Local AI
- Ollama · on-device
- Model server
- Detect · pull · test
- Provider APIs
- Explicit per turn
Local AI
Own the model, server, and provider boundary
Kairos detects a loopback Ollama server, guides setup, pulls selected models, checks service health, and runs a real 32K-context test. Local inference is the default; OpenAI, Anthropic, and CLI providers remain modular choices with a per-turn egress preview and no silent fallback.
How it works
Route before reading
Every turn follows an explicit path from intent to evidence. No note body is retrieved, sent, or changed before the relevant policy decision is made.
- 01RouteChoose the approved brain for the question.
- 02RetrieveAssemble a small, bounded context pack.
- 03ChooseRun local inference or preview an API/CLI hand-off.
- 04AnswerReturn the response with visible source citations.
- 05ConfirmRequire approval before any Markdown write.
Product
Kairos is a macOS companion for working across multiple personal knowledge systems and AI runtimes. It can run models locally, control when external providers are used, keep the original notes authoritative, and return a small cited context pack rather than silently searching everything.
Why it exists
Research, projects, and everyday work have different confidentiality and access rules. A general-purpose chat interface can lose provenance or cross those boundaries. Kairos makes routing, retrieval, provider choice, and writes explicit parts of the interaction.
Local AI and provider control
The local model path is a managed product capability, not a demo toggle. Kairos works with the loopback Ollama service, recommends models against available hardware, manages model downloads and health checks, and verifies the selected model at the intended context window. External APIs and installed AI CLIs are separate adapters that require an explicit destination and content preview for each hand-off.
Working system
- Explicit routing across registered knowledge systems
- Bounded retrieved context with visible source citations
- Local Ollama server detection, model pulls, health checks, and context tests
- Consented provider API and CLI hand-offs with a per-turn preview
- Approval-gated Markdown create and edit proposals
- A metadata-only atlas for navigating connected knowledge
Build
- Tauri and Rust for the desktop application and policy layer
- React and TypeScript for the macOS interface
- Ollama loopback model server and hardware-aware model catalogue
- OpenAI, Anthropic, Codex CLI, and Claude Code adapters behind explicit consent
Maturity
What exists now
| Area | State | Evidence |
|---|---|---|
| Routing and policy core | Working | Registered brains, scoped retrieval, and explicit access rules. |
| Cited context | Working | Bounded excerpts retain their source provenance. |
| Local AI runtime | Working | Loopback Ollama detection, setup guidance, model pulls, health checks, and 32K tests. |
| Provider and API control | Working | Explicit OpenAI, Anthropic, and CLI adapters with per-turn egress previews and no silent fallback. |
| Signed public release | Not yet | Packaging, notarisation, onboarding, and support remain unfinished. |
Interested in Kairos or local AI?
Discuss private AI workflows, local model serving, provider boundaries, or the current Kairos developer alpha.
Explore the other independent products in Tharm Lab.
All products