convez.ai
MCP ServerAutomaçãoAvançado
lionwashington

KA Toolkit — Assistente de Conhecimento Local

lionwashington/ka-toolkit

Capture LLM conversations em KB local com MCP servers e integração Telegram/Lark

Ver no GitHub
InicialPluginScore · de 0 a 100

Toolkit local-first que captura conversas com LLMs em um repositório Markdown compatível com Obsidian, oferecendo MCP servers para busca/leitura de tópicos e daemons para Telegram/Lark. Funciona como camada de integração sobre runtimes de agentes (Claude Code, Gemini CLI) sem substituí-los, mantendo soberania total dos dados locais.

Instalação

$npx convez-ai lionwashington/ka-toolkit

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 dia
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Construir uma base de conhecimento pessoal a partir de interações com agentes de IA
  • Integrar sessões de agentes com canais do Telegram ou Lark para orquestração em tempo real
  • Pesquisar e recuperar tópicos de uma KB Markdown sem depender de serviços em nuvem

Veredito

Projeto ambicioso e bem-arquitetado para quem quer agência local com soberania de dados. Exige TypeScript, pnpm e compreensão de MCP + Linux/tmux; ideal para engenheiros que querem orquestração distribuída entre múltiplos agentes LLM.

README

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

ka-toolkit

A local-first personal knowledge-assistant toolkit built on top of an agent runtime (Claude Code today). It captures your LLM conversations, distills them into an Obsidian-compatible Markdown knowledge base, and ships a set of MCP servers and a chat-channel daemon so you can drive agent sessions from Telegram or Lark.

Design principle — local-first & runtime-agnostic. Your knowledge base is plain Markdown on your own disk. The toolkit is hosted on an agent runtime (CC / Codex / Gemini CLI) rather than replacing it, reuses your existing subscriptions, and never trades local data sovereignty for a cloud service.


What's in here

A pnpm + sh monorepo, organized by four functional parts (capability / orchestration / communication / scheduling) plus shared / config / tests:

PathPart — what it does
kb/coreCapability. The knowledge pipeline: capture, distill, retrieval, knowledge-store, daily-log / topic splitters (CJK-aware)
kb/mcp-serverKB MCP server: kb_search / kb_read_topic / kb_list_topics / kb_status
kb/skill, kb/skillsAgent skills: /kb, daily-brief, mail, calendar, and more
kb/tools/*Domain MCP servers: hkprop-mcp (HK rentals · 28Hse + Centanet), ibkr-mcp (IBKR P&L), market-mcp (equity/crypto quotes), mcp-opennutrition (food lookup)
kb/adapter-ccClaude Code runtime adapter (capture hook, scheduler, install)
kb/opsBackground distill worker scripts
channels/coreCommunication. Platform-independent MCP-over-HTTP daemon kernel (sessions, routing, dispatch, re-adopt self-heal)
channels/telegram, channels/larkTelegram / Lark platform adapters; channels/ops/daemon.sh is the ka channel CLI
workshop/opsOrchestration. ka workshop (tmux panes / CCs) + start-pane / wait-ready / tmux & yaml helpers
cron/opsScheduling. ka cron + the run / plist / crontab backend
shared/ops, shared/bin/kaCross-cutting: the ka dispatcher + common.sh / doctor / status / help
config/, tests/Bundled config templates; the Docker test harness

How it fits together

  • Capture → distill → KB. Conversations are captured and distilled into ~/.knowledge-assistant/ as Obsidian-compatible Markdown (daily logs auto-split when they grow past a threshold; topics get hub + sub-topic files).
  • MCP servers expose the KB and the domain tools to any MCP-capable agent.
  • Channel daemon (channel-core + a platform adapter) runs as a standalone MCP-over-HTTP service. Each agent session connects to it, so you can send a message in Telegram / a Lark group and the corresponding session receives it and can reply.
  • ka workshop brings up multiple agent sessions, each in its own tmux pane / cwd / channel, all bound to the daemon.
  • Design / runtime separation. This repo is pure design-time source. install.sh builds and copies the artifacts into ~/.knowledge-assistant/; at runtime everything runs the deployed copy and never points back at the repo.

Getting started

pnpm install
pnpm rebuild esbuild   # pnpm 10+ gates native build scripts; daemon/MCP bundling needs esbuild
pnpm build

Configuration and secrets live in ~/.knowledge-assistant/ (e.g. config.yaml, secrets.yaml) and are never committed — see config/secrets.example.yaml for the shape.

Tests

pnpm test                                   # JS/TS packages
cd kb/tools/hkprop-mcp && uv run pytest      # Python MCP

Status & scope

This started as a personal tool and is shared as-is. The Claude Code runtime adapter is the only one implemented; Codex / Gemini CLI are reserved seams. Expect rough edges and opinions baked in — it is optimized for one person's workflow, not as a turnkey product.

License

Apache License 2.0.

Ver o original no GitHub

Bom para

Relacionados