KA Toolkit — Assistente de Conhecimento Local
lionwashington/ka-toolkit
Capture LLM conversations em KB local com MCP servers e integração Telegram/Lark
Ver no GitHubToolkit 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-toolkit1 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
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:
| Path | Part — what it does |
|---|---|
kb/core | Capability. The knowledge pipeline: capture, distill, retrieval, knowledge-store, daily-log / topic splitters (CJK-aware) |
kb/mcp-server | KB MCP server: kb_search / kb_read_topic / kb_list_topics / kb_status |
kb/skill, kb/skills | Agent 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-cc | Claude Code runtime adapter (capture hook, scheduler, install) |
kb/ops | Background distill worker scripts |
channels/core | Communication. Platform-independent MCP-over-HTTP daemon kernel (sessions, routing, dispatch, re-adopt self-heal) |
channels/telegram, channels/lark | Telegram / Lark platform adapters; channels/ops/daemon.sh is the ka channel CLI |
workshop/ops | Orchestration. ka workshop (tmux panes / CCs) + start-pane / wait-ready / tmux & yaml helpers |
cron/ops | Scheduling. ka cron + the run / plist / crontab backend |
shared/ops, shared/bin/ka | Cross-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 workshopbrings 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.shbuilds 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
- macOS install: see
docs/INSTALL.md - Ubuntu / WSL2 with a Lark daemon: see
docs/INSTALL_UBUNTU.md - Architecture overview:
docs/ARCHITECTURE.md kaCLI usage:docs/KA_CLI_USAGE.md
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.