The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Kogcat MCP listing page.
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 Claude Code and Codex. Before you act on an AI answer, it surfaces the counterexamples, the boundaries, the blind spots — drawn from a knowledge base that lives on your machine. It won't replace your model. It won't slow you down. The call stays yours.
"I read 30 minutes a day but nothing sticks. Should I 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. Your knowledge base holds a claim you marked high-confidence: the "I get it" feeling while re-reading is the least reliable signal of real recall. So try this once — finish a section, close the book, write what you remember. Then compare it to what you thought you had.
Built for judgment, not lookup. KogCat speaks up for the calls that cost you when they're wrong — decisions, tradeoffs, comparisons, critiques, strategy. It stays quiet for the rest: lookups, definitions, code, summaries, translation. And even on a judgment call, it adds a note only when it sees something the model didn't.
/kogcat:query <question> puts your knowledge base first: a conclusion, the conditions that change it, a next step.Claude Code
Codex
After installing, fully quit and reopen Claude Code (or restart Codex). The first-run download only starts on the next fresh session — the install command alone won't begin it.
On first launch, KogCat quietly downloads its engine (~40 MB) and embedding model (~90 MB). Once. About a minute on a good connection. Keep working while it does — run /kogcat:status anytime to watch each piece come online.
KogCat's calibration engine runs as a local sidecar; the Claude Code / Codex plugin is just one client of it. Any MCP-capable tool — Cursor, Cline, Zed, VS Code, Claude Desktop — can use the same engine through a standalone stdio MCP server.
Add this to your client's MCP config (field names vary slightly by client; most use an mcpServers map):
Requires uv, on macOS (Apple Silicon) or Windows x86_64 (same engine, same platforms as below). On first run it downloads the engine + embedding model and registers a background sidecar — the same one-time setup as the plugin, shared by every client on the machine. It exposes the knowledge-base tools (search, node, edges, calibrate, calibrate_review, and the memory_* family) for your model to call.
What you give up vs. the plugin. The Claude Code / Codex plugin adds two host conveniences a generic MCP client has no hook for: calibration that fires automatically on judgment questions, and a memory index injected into context at session start. With a standalone server your model reaches the same knowledge base, but it's the model that decides to call those tools — or you ask it to "use KogCat" — rather than a hook firing them for you. You never type the tool names (search, calibrate_review, memory_*); they're the model's to call.
No account. No subscription. No one else holding your knowledge.
| Command | What it does |
|---|---|
/kogcat:query <question> | A knowledge-base-first answer: conclusion, conditions, next step. |
/kogcat:status | A read-only local check. Reach for it if first launch seems stuck. |
/kogcat:memory-consolidate | Review and tidy saved memories — every change is yours to confirm. |
Automatic calibration needs no command.
PATH — already there on macOS; on Windows, install it yourself (tick Add python.exe to PATH)FSL-1.1-MIT — see LICENSE. Converts to MIT two years after each release.