convez.ai
AgenteProdutividadeIntermediário
hesorchen

muselab

hesorchen/muselab

Assistente IA privado que conhece seu arquivo pessoal inteiro

Ver no GitHub
BomPluginScore · de 0 a 100

Agent de IA self-hosted que funciona como um assistente inteligente para seu arquivo pessoal de documentos, mantendo tudo privado na sua máquina. Baseado no mesmo loop de agente do Claude Code, permite que Muse leia e trabalhe com arquivos sensíveis (PDFs de saúde, planilhas financeiras, notas) sem enviar dados para a nuvem, gerando relatórios HTML e documentos renderizados em tempo real.

Instalação

$npx convez-ai hesorchen/muselab

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
66
Forks
8
Acompanhando
0
Saúde
85%
Último commit
há 3 dias
Versão
v2.1.0
Issues abertas
9
Contribuidores
1

Casos de uso

  • Análise comparativa de documentos pessoais (saúde, finanças) com geração de relatórios
  • Revisão e anotação de artigos e papers armazenados localmente
  • Cruzamento de informações entre múltiplos domínios (saúde + seguro + finanças) para decisões integradas
  • Criação de documentos e relatórios HTML/Markdown processados em tempo real

Veredito

Ferramenta ambiciosa e bem-pensada que resolve um problema real: manter contexto pessoal sensível privado enquanto aproveita capacidades de IA avançadas. A arquitetura self-hosted é sólida e o suporte a múltiplos provedores de modelo (Claude, DeepSeek, GLM, etc.) oferece flexibilidade; recomendado para usuários que precisam lidar com documentos confidenciais e estão dispostos a fazer deploy próprio.

README

A tradução para o português deste README ainda está na fila.

muselab

CI License: MIT Self-hosted Container Ask DeepWiki 中文

muselab is to your life's files what Claude Code is to your codebase.

Models get replaced every month. Your context doesn't — it compounds. The files you'd never hand to a SaaS — checkup PDFs, budget spreadsheets, papers you've read, half-written notes — are exactly what AI should be reading. muselab is a self-hosted AI workspace: the archive stays on your own disk, and Muse — built on the same agent loop that powers Claude Code — works on it directly. The only thing that ever leaves your machine is the request sent to the model you picked.

  • 🔐 Private, so you hold nothing back. No SaaS account, no cloud copy — which is why health, money and work can finally live in one archive. Muse reads across all of it at once, and gives the kind of advice no single-domain view can produce.

  • 📈 Your context compounds. Whole files enter the context exactly as written — no vectorizing, no chunking, no retrieval index. Every new model generation is a free upgrade to your assistant, because the archive it lands on is already there, and growing. Eight providers, one click apart: Claude / DeepSeek / GLM / MiniMax / Kimi / Qwen / MiMo / ERNIE — swap the engine anytime, the asset stays yours. Reuse the Claude subscription you already pay for, or run on dirt-cheap API models.

  • 📄 Deliverables, not chat bubbles. Muse writes HTML reports and Markdown docs that render live in the preview pane as it types — no plugins, no setup. A paper becomes an annotated reading page; a folder of statements becomes a charted report.

  • 📱 An agent in your pocket. Claude Code lives in a terminal. A muselab task started at your desk can be steered from your phone on the way out — install as a PWA, get a push when a long run finishes.

What a session looks like

"Compare this new checkup PDF with last year's, and turn the changes into a one-page HTML trend report."

Muse greps health/ for both PDFs, reads them whole, extracts the numbers, and writes a single-file HTML report with charts — rendered live in the preview pane. Then you follow up:

"Now check the insurance policies in money/ — do any of these changes leave a gap?"

That's the crossing: two domains in one context is what turns answers into actions. And on your way out, open the same session on your phone and keep going.

🌐 More scene demos on the muselab promo page.

Why not just ChatGPT?

What you use todayWhere it stopsmuselab
ChatGPT / Claude.aiFiles re-uploaded per chat, memory is a black box, sensitive archives stay outThe archive lives on your disk, readable in full
Claude CodeThe strongest agent loop — but born in a terminal, built for codeThe same loop, pointed at your life's files, in a browser and on your phone
RAG document chatChunk-and-retrieve loses meaning across documentsWhole files in context, zero loss

Full comparison (Open WebUI, LobeChat, AnythingLLM, claudecodeui …): How it compares.

Small things, done right

  • Queue without losing a word — keep typing while Muse works; a server-side FIFO queue runs each message in turn
  • Scheduled tasks — daily / weekly / monthly / once; missed runs catch up after downtime; results land in a bell drawer and push to your phone
  • Session forking — branch from any message, rewrite and re-run
  • Restart recovery — sessions and queued messages come back exactly as they were
  • A real file tree — drag-and-drop upload, search, inline rename, drag-to-trash
  • Three themes × accent picker — light / dark / eye-care, with your choice of accent color
  • Bilingual UI — English / 中文, one click
  • No build step — edit a frontend file, refresh the browser

Install

Prerequisites: git, curl (bundled on Linux / macOS; on WSL2 run sudo apt install git curl).

One-line (Linux + macOS + WSL2) — installs uv, clones into ~/muselab, then runs the platform installer (which auto-installs Node LTS + the Anthropic claude CLI and registers the service):

curl -fsSL https://raw.githubusercontent.com/hesorchen/muselab/main/scripts/quick-install.sh | bash

Windows users: install via WSL2 (see Quick start).

Unattended — for CI / Docker / demo recording. Takes every default (random token, port 8765, ~/muselab-archive) and skips every prompt:

curl -fsSL https://raw.githubusercontent.com/hesorchen/muselab/main/scripts/quick-install.sh | MUSELAB_NONINTERACTIVE=1 bash

Manual — if you'd rather see every step:

git clone https://github.com/hesorchen/muselab && cd muselab
bash scripts/install-linux.sh    # or install-macos.sh

Open http://localhost:8765, paste the token from .env. If the installer reported "claude CLI is installed but not logged in", run claude login once to enable Anthropic models.

For prerequisites, Docker, dev mode and per-OS detail, see Quick start.

Docs

📚 Full documentation index

Status

v1.0 — first stable release. If compounding context resonates with you, a ⭐ helps more people find it — and the best day to start your archive is today. PRs are welcome — see CONTRIBUTING.md. The roadmap and known issues are tracked on GitHub Issues.

MIT

Ver o original no GitHub

Bom para

Relacionados