Klava

A personal AI assistant for founders, BD, sales, and CEOs - built on top of Claude Code.

If you're a developer, Claude Code is already a superpower. If your work is mostly deals, meetings, research, and operations - the CLI alone doesn't cut it. Klava is what I built for myself over 9 months to make Claude actually useful for non-code work: a dashboard, a task queue, a knowledge base, proactive jobs, and 40+ skills for real workflows.

Full story and design rationale: blog post (ToDo: add link).

Install

Two steps, plus a short setup wizard.

git clone https://github.com/VCasecnikovs/klava.git
cd klava
./setup.sh

setup.sh handles submodules, Python venv, Node build for the dashboard, LaunchAgents, config templates, Claude Code authentication, and opens the dashboard at http://localhost:18788/dashboard. On first run it walks you through a short wizard: your name, timezone, Telegram bot token (optional), Obsidian vault path.

That's it. Cron jobs and the Telegram bot stay off until you explicitly enable them from the Settings tab.

What you get

A dashboard with chat, tasks, and proactive surfaces

The primary interface. Left side - chat with Klava. Right side - supporting tabs: Deck, Tasks, Health, Views, Skills.

Markdown, tables, and SVG render natively. You can quote any part of an agent message like in WhatsApp.

A Deck of cards for tasks

Every task - proposal, action, result - shows up as a card. One at a time. Done, Skip, Snooze, or open a Session to work on it with Klava.

When Klava finishes a delegated task, she drops a Result card back on the Deck.

Views for interactive reports

Some things don't fit in a chat - design prototypes, long research, comparison matrices. Klava generates them as interactive HTML pages you can open as views.

Vadimgest for reacting to the world

Vadimgest is a separate submodule that collects updates from 19 sources (messages, emails, calendar, browser history, etc.) and turns them into LLM-readable feeds. A heartbeat job reads it every 30 minutes and decides what to do: create tasks, update the knowledge base, or act without you.

How it's built

Component	Purpose
Webhook server (gateway/webhook-server.py)	Dashboard, API, real-time chat via WebSocket
CRON scheduler (gateway/cron-scheduler.py)	Heartbeat, reflection, health checks
Telegram bot (gateway/tg-bot.py)	Proactive notifications and mobile access
Watchdog (gateway/cron-watchdog.py)	Monitors daemons, restarts stale processes
Klava consumer (tasks/consumer.py)	Executes delegated tasks from the Klava queue

Scheduled jobs:

• Heartbeat (every 30 min) - reads Vadimgest, triages new data, creates tasks, updates knowledge base.
• Reflection (nightly) - cleans up Inbox, cross-links entities, dedupes tasks, detects silence on deals.
• Klava consumer (every 5 min) - picks one task off the async queue, runs it, writes a Result card back to the Deck.

Skills

Skills are markdown files in .claude/skills/*/SKILL.md that teach Klava specific workflows - how to draft a cold email, how to record a deal, how to generate a pitch deck. Claude can create and edit its own skills.

40+ shipped skills including: autoplan, autoresearch, cold-email, comms, heartbeat, reflection, meeting-processing, language-coach, voice, html-view, pdf, xlsx, docx, pptx, codex-cli, gemini-cli, mcp-builder, skill-creator.

Create your own:

mkdir .claude/skills/my-skill
cat > .claude/skills/my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: What this skill does in one line
user_invocable: true
---

# My Skill

Instructions for Claude go here...
EOF

Skills hot-reload - no restart needed. Use /my-skill in chat to invoke.

Personal vs shipped skills

For skills with private data (API keys, personal workflows), use .claude/skills/personal/ - it's gitignored. Link them in with:

./scripts/link-personal-skills.sh

Configuration

gateway/config.yaml

Central config. Edit via the Settings tab or directly. Key sections: identity, you, telegram, models, cron, paths.

cron/jobs.json

Scheduled job definitions. Each job has schedule (cron or interval), execution (Claude session or bash), retry, catch_up. See cron/jobs.json.example.

Memory tiers

1. Raw data - Vadimgest (append-only JSONL, 19 sources)
2. Structured knowledge - Obsidian vault (People, Organizations, Deals, Topics) + skills
3. CLAUDE.md + MEMORY.md - high-frequency system instructions loaded every session

CLAUDE.md is tracked and generic. MEMORY.md is gitignored and personal.

Prerequisites

• macOS (LaunchAgents for daemon management - Linux support via systemd is welcome as a PR)
• Python 3.11+
• Node.js 18+ (dashboard build)
• Claude Code installed and authenticated
• Obsidian vault (optional, knowledge graph)
• Telegram bot token (optional, proactive notifications)

Directory structure

.claude/
  CLAUDE.md              # Generic system instructions (tracked)
  MEMORY.md              # Personal overlay (gitignored)
  skills/                # Generic skills (tracked)
    personal/            # Personal skills (gitignored)
  pipelines/             # Workflow state machines

gateway/
  webhook-server.py      # Dashboard + API
  cron-scheduler.py      # Job scheduler
  tg-bot.py              # Telegram bot
  cron-watchdog.py       # Health monitor
  lib/                   # Shared libraries
  hooks/                 # Event hooks

tools/
  dashboard/react-app/   # React + TypeScript + Vite

cron/
  jobs.json              # Job definitions

tasks/
  queue.py               # Klava task queue
  consumer.py            # Queue executor

vadimgest/               # Data-lake submodule (optional, private)

Contributing

Started as a personal project, now being open-sourced incrementally. PRs welcome on:

• Linux support (systemd instead of LaunchAgents)
• Alternative notification channels (Discord, Slack, email)
• New generic skills
• Dashboard improvements
• Documentation

License

Apache License 2.0