KogCat

AI makes the answer smoother. KogCat makes the judgment sound.

English | 中文 · Website: https://www.kogcat.com

A local-first judgment calibration layer for Obsidian. Select a line, a section, or a whole note. KogCat reads it against a structured knowledge base of cognitive biases, logic fallacies, and decision frameworks — and hands back a few calibration points to sharpen the judgment. Your markdown stays exactly as you wrote it.

See the difference

You wrote:

I read 30 minutes a day but feel like I'm forgetting everything. I should take more detailed notes.

Plain AI — Try the Cornell method, highlight key passages, add Anki for spaced repetition.

KogCat — More notes will likely make it worse. The bottleneck isn't capture. It's retrieval. Research is consistent: close the book, write what you remember — even imperfectly — and it sticks better than any note format applied while reading. Try one session that way. Then weigh what you actually kept against what you thought you had.

How it works

Open the review panel — from the ribbon, the command palette, or a right-click — on a selection, a paragraph, or the whole note. The local engine pulls a recall pool from its calibration base. Connected a chat model? It distills the pool to the few points that matter. Without one, a built-in fallback still works.

Privacy & network

KogCat has no server of its own. Exactly two things touch the network:

1. The calibration engine (om-core). On first launch, KogCat downloads the binary for your platform from the public release channel (GitHub Releases, with an Alibaba OSS mirror) and verifies it against a sha256 manifest before it runs. It's cached outside your vault and reused. KogCat registers it as a local background service — a com.kogcat.om LaunchAgent on macOS, an OmCore Task Scheduler task on Windows — and talks to it over a local socket only. No inbound port. Prefer your own binary? Point to it in Settings → KogCat → engine path, or set OM_ALLOW_DIRECT_SPAWN=1 to skip the service entirely.
2. Your LLM provider. Called only when you consent to review refinement — bring your own key. Point it at a local model (Ollama / LM Studio) to stay fully offline.

No telemetry. No analytics. No account. Your vault is never uploaded — only the text you hand KogCat to review is read, and it never writes proprietary syntax into your notes.

Maintenance. Actively maintained. The calibration base refreshes through the engine — nothing for you to install or update. If we ever stop, you'll hear it in advance, with a clean data-export path.

Uninstall. Remove the plugin and your vault is untouched. Two things live outside it: the background service and the cached engine. Stop and delete the service (the LaunchAgent or OmCore task), then delete the engine under your OS application-support folder.

Requirements

• Obsidian 1.0.0+ (desktop only)
• macOS (Apple Silicon) or Windows x64 — Intel Mac and Linux not yet supported
• An API key for one supported provider: Anthropic, OpenAI, Google Gemini, xAI, DeepSeek, Mistral, Perplexity, OpenRouter, Azure OpenAI, or any OpenAI-compatible endpoint. Local providers (Ollama / LM Studio) need no key.

Install

1. Install KogCat from the Obsidian community plugin browser, or drop a release build into <vault>/.obsidian/plugins/kogcat/, then enable it.
2. Open Settings → KogCat and enter your provider API key.
3. Trigger a review. On first use the engine downloads automatically, verifies, and starts; later launches reuse the cache.

Usage

Select text, or rest the cursor in a paragraph. Open KogCat from the ribbon, command palette, or right-click menu. Then revise the original passage against the points in the panel.

Acknowledgments

Forked from Smart Composer by Heesu Suh. KogCat keeps its multi-provider LLM client and settings scaffolding; the chat, RAG, MCP, and prompt-template subsystems were removed.

License

FSL-1.1-MIT — converts to MIT two years after each release. Portions derive from Smart Composer (Copyright (c) 2024 Heesu Suh), originally MIT-licensed; see Acknowledgments above.