{
  "markdown": "# KogCat\n\n> AI makes the answer smoother. KogCat makes the judgment sound.\n\n[English](./README.md) | [中文](./README.zh-CN.md) · **Website:** <https://www.kogcat.com>\n\nA 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.\n\n---\n\n## See the difference\n\n*\"I read 30 minutes a day but nothing sticks. Should I take more detailed notes?\"*\n\n**Plain AI** — Try the Cornell method, highlight key passages, add Anki for spaced repetition.\n\n**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.\n\n---\n\n## When it speaks up\n\nBuilt 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.\n\n- **Automatic.** Ask a judgment question in conversation. A note appears only when there's something worth saying — and your original answer is never touched.\n- **On demand.** `/kogcat:query <question>` puts your knowledge base first: a conclusion, the conditions that change it, a next step.\n\n---\n\n## Install\n\n**Claude Code**\n\n```\n/plugin marketplace add KogCat/cc-kogcat\n/plugin install kogcat\n```\n\n**Codex**\n\n```\ncodex plugin marketplace add KogCat/cc-kogcat\ncodex plugin add kogcat@kogcat\n```\n\n**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.\n\nOn 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.\n\n---\n\n## Use it from any MCP client\n\nKogCat'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.\n\nAdd this to your client's MCP config (field names vary slightly by client; most use an `mcpServers` map):\n\n```json\n{\n  \"mcpServers\": {\n    \"kogcat\": {\n      \"command\": \"uvx\",\n      \"args\": [\"kogcat-mcp\"]\n    }\n  }\n}\n```\n\nRequires [uv](https://docs.astral.sh/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.\n\n**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.\n\n---\n\n## Privacy\n\n- **Your knowledge base** stays on your machine. KogCat reads only the folder you point it at.\n- **Your conversation** goes to the same Claude or Codex you already use. Nothing extra, nowhere else.\n- **Calibration** happens in a local process. The results never leave.\n- **The engine** comes from a public release channel, checked against a sha256 manifest before it ever runs.\n\nNo account. No subscription. No one else holding your knowledge.\n\n---\n\n## Commands\n\n| Command | What it does |\n|---|---|\n| `/kogcat:query <question>` | A knowledge-base-first answer: conclusion, conditions, next step. |\n| `/kogcat:status` | A read-only local check. Reach for it if first launch seems stuck. |\n| `/kogcat:memory-consolidate` | Review and tidy saved memories — every change is yours to confirm. |\n\nAutomatic calibration needs no command.\n\n---\n\n## Requirements\n\n- Claude Code or Codex\n- macOS (Apple Silicon) or Windows x86_64 — Intel Mac and Linux not yet supported\n- Python 3 on `PATH` — already there on macOS; on Windows, install it yourself (tick *Add python.exe to PATH*)\n\n---\n\n## License\n\nFSL-1.1-MIT — see [LICENSE](./LICENSE). Converts to MIT two years after each release.\n",
  "bytes": 4997,
  "sha": "f3c0513af370ff49acfd354dccc3471635bb9314b4233277733233d612fe1730",
  "repo_slug": "kogcat/cc-kogcat",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_com_kogcat_kogcat_mcp_8c795d6c/readme"
}