neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
翻译准备中
尚未检测到更新。GitHub 发布版本后会显示在这里。
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权,自行管理介绍,无需管理员批准。
登录并生成验证文件,在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件,即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录,将前往登录页面。登录后请回到此处,再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify,即可验证所有权并自行管理介绍。
一分钟完成验证MCP server for Lupa — let your AI agent search your Mac's documents by content (PDF, DOCX, HWP, OCR'd scans). Korean-aware, ~30ms, fully local.
翻译准备中
取自仓库 README
正在加载完整 README。 以下为已保存的摘录。
Let your AI agent search the documents on your Mac — by what's written inside them.
An MCP server for Lupa, a macOS file search app with its own index.
Your agent asks Lupa; Lupa answers from a local index in ~30 ms. Nothing touches the network.
Searches file names and document contents: PDF, Word, Excel, PowerPoint, plain text,
Hangul HWP/HWPX, and scanned PDFs read with OCR.
That last row is the point. Your agent gets document contents without you granting it
access to your disk.查看完整 README ↗Let your AI agent search the documents on your Mac — by what's written inside them.
An MCP server for Lupa, a macOS file search app with its own index. Your agent asks Lupa; Lupa answers from a local index in ~30 ms. Nothing touches the network.
"find the contract that mentions the penalty clause"
↓
lupa_search → ranked paths + matching snippets
lupa_read → the document's text, without opening the file
Searches file names and document contents: PDF, Word, Excel, PowerPoint, plain text, Hangul HWP/HWPX, and scanned PDFs read with OCR.
find / grep / mdfind | lupa-mcp | |
|---|---|---|
| Searches inside PDF, DOCX, HWP | no | yes |
| Scanned documents (OCR) | no | yes |
| Korean partial words and particles | poorly | yes — 보호정책 finds 정보보호정책서.hwp |
| Speed on a large disk | seconds | ~30 ms (indexed) |
| Agent needs file system access | yes | no — Lupa reads the files, not your agent |
That last row is the point. Your agent gets document contents without you granting it access to your disk.
The Lupa app must be installed — this server reads the index that Lupa builds; it does not index anything itself.
macOS only.
Claude Code
claude mcp add lupa -- npx -y lupa-mcp
Claude Desktop / Cursor / any MCP client — add to the client's MCP config:
{
"mcpServers": {
"lupa": { "command": "/bin/zsh", "args": ["-lc", "npx -y lupa-mcp"] }
}
}
That's it. No API key, no account, no configuration.
Why
/bin/zsh -lc? GUI apps don't inherit your shell's PATH, so a bare"command": "npx"fails to start withenv: npx: No such file or directory(measured on Claude Desktop and Aside). A login shell findsnpxwherever it lives — Homebrew, nvm, or a system install. In a terminal-based client (Claude Code, Codex) plainnpx -y lupa-mcpis fine.
Agents: don't run
lupa-searchthrough your shell. Sandboxed agent shells (Aside, Codex) kill the App Store-signed CLI at launch — exit code 133 or 134. Register this MCP server instead; it starts outside that sandbox. Full guide: https://lupa.kr/agents.html
lupa_search| argument | default | meaning |
|---|---|---|
query | — | see syntax below |
limit | 10 | results to return (max 50) |
names_only | false | file names only, ignore contents (faster) |
snippet_chars | 160 | truncate each snippet (0 = none) |
lupa_readReturns the text Lupa already extracted from a file — cheaper than reading the original, and it works for formats your agent cannot parse. Requires Lupa 2.0 or later.
| argument | default | meaning |
|---|---|---|
path | — | absolute path, as returned by lupa_search |
max_chars | 8000 | cap the returned text (0 = no limit) |
word1 word2 files matching ALL words rank first
n:word file name only (name:, 이름:, 文件名:)
c:word document content only (content:, 내용:, 内容:)
f:word folder path contains (폴더:, 文件夹:)
-word exclude
"exact phrase" literal match
.pdf *.mp3 report* extension / wildcard
kind:document document · pdf · hwp · image · video · audio · code · archive · app
size:>10mb size:<1mb
date:today date:week · date:month · date:7d · date:2026-06
Korean and Chinese prefixes work too (종류: 크기: 날짜: / 种类: 大小: 日期:).
Korean is handled properly — partial words inside compounds and attached particles both match, which is what Spotlight gets wrong.
Every tool call is recorded on your Mac so you can see what an agent searched for and which files it was shown:
~/Library/Logs/Lupa/mcp-audit.jsonl
Each line has the time, the calling client (e.g. Claude Code, Cursor), the tool, the query or path, the paths returned, and whether it succeeded. Snippets and document text are not logged. The file is readable only by you (0600) and never leaves the Mac.
Lines are hash-chained: each entry includes the hash of the previous one, so editing or removing a line is detected. This makes the log tamper-evident, not tamper-proof — deleting the whole file and starting over is still possible.
npx lupa-mcp audit # last 20 calls
npx lupa-mcp audit -n 100
npx lupa-mcp audit --verify # exit 0 if the chain is intact, 1 if a line was changed or removed
Turn it off with LUPA_MCP_AUDIT=off, or move it with LUPA_MCP_AUDIT_PATH=/path/to/file.jsonl.
Everything runs on your Mac. Lupa has no network code at all, and this server only launches Lupa's bundled command-line tool and passes the results to your agent. Nothing is uploaded, and there is no telemetry.
lupa_read truncates long files.LUPA_SEARCH_PATH to the lupa-search
binary inside the app bundle.| message | fix |
|---|---|
| "Lupa를 찾을 수 없습니다" / not found | Install Lupa from the Mac App Store |
| "검색 인덱스가 없습니다" | Open Lupa, add folders, wait for indexing |
| "본문 읽기를 지원하지 않습니다" | Update Lupa to 2.0 or later |
| No results for a file you know exists | Its folder may not be registered in Lupa |
MIT