English | Русский
A self-hosted personal AI agent: a markdown wiki, a Things-style task manager, and a universal assistant in Telegram — all driven by a single agent built on the Claude Agent SDK. Runs on your Claude subscription (OAuth via Claude CLI), no API keys required.
- Tasks — Things-style mechanics: Inbox / Today / Upcoming / Someday, projects and areas, tags, deadlines, checklists, repeating tasks, logbook, drag-and-drop, hotkeys, PWA. Live sync over SSE — changes made from Telegram or by scheduled jobs appear on screen by themselves.
- Wiki — a personal knowledge base of markdown files. The agent reads and writes pages, cross-links them, and keeps things tidy.
- Two UI languages — English and Russian: switchable in Tasks settings; the wiki follows the browser language.
- Assistant everywhere — web chat in both UIs plus a Telegram bot sharing one session: whatever you discussed on the web, it remembers in Telegram. Voice messages via ASR. Replies stream in Telegram through the native
sendMessageDraft. - Scheduling — "remind me in 20 minutes", "send my tasks every weekday at 8:30": the agent creates cron jobs itself. Every run sees the outputs of previous runs (no repeating itself), stays quiet when there is nothing new (
[SILENT]), and stops the job once the tracked event is over ([FINAL]). - Memory & self-improvement — long-term memory about the user (survives session resets), self-authored skills, a background reviewer after every turn (decides what to persist), a weekly skill-library curator, and a session freshness window.
- Subagents — researcher (web research) and librarian (wiki reorganization) via Task.
| Dark theme & palettes | Project with logbook |
|---|---|
 |
 |
backend/ FastAPI + claude-agent-sdk (single process)
app/agent.py sessions, streaming, memory snapshot, freshness window
app/scheduler.py cron ticker (60s), [SILENT]/[FINAL], run history
app/reviewer.py background post-turn reviewer (memory/skills)
app/telegram.py bot: long polling, draft streaming, /status
app/tasks_*.py Things mechanics on SQLite (+SSE)
agent_skills/ the agent's domain skills (wiki/tasks)
frontend-wiki/ React: three panes, markdown, chat
frontend-tasks/ React: tasks, dnd-kit, themes & palettes, chat
Storage is files and SQLite on a volume: content/ (markdown wiki) and data/ (tasks, cron, memory, skills, session). Neither is in the repository — that's personal data.
You need Docker and an authenticated Claude CLI (the agent uses its OAuth credentials from ~/.claude).
git clone https://github.com/0717376/bender && cd bender
cat > .env <<'ENV'
WIKI_PASSWORD=pick-a-password
CLAUDE_MODEL=sonnet
# Telegram (optional): bot token and your chat id
TELEGRAM_BOT_TOKEN=
TELEGRAM_ALLOWED_IDS=
ENV
docker compose up -d --build- Tasks: http://localhost:8851
- Wiki: http://localhost:8842
The first message to the Telegram bot will tell you your ID — put it into TELEGRAM_ALLOWED_IDS and restart the backend.
| Variable | Default | Purpose |
|---|---|---|
WIKI_PASSWORD |
— | web UI password (required) |
CLAUDE_MODEL |
sonnet |
agent model (sonnet/opus/haiku) |
TELEGRAM_BOT_TOKEN |
— | bot token; empty disables the bot |
TELEGRAM_ALLOWED_IDS |
— | comma-separated chat ids |
ASR_UPSTREAM |
— | speech-to-text service URL for voice messages |
ASR_MODEL |
gigaam-rnnt |
model_id passed to the ASR service |
SESSION_FRESH_HOURS |
6 |
idle time after which a fresh session starts |
REVIEWER_ENABLED / REVIEWER_MODEL |
1 / sonnet |
background memory/skills reviewer |
CURATOR_ENABLED / CURATOR_INTERVAL_HOURS |
1 / 168 |
skill-library curator |
CLAUDE_DIR / CLAUDE_JSON |
~/.claude / ~/.claude.json |
Claude CLI credentials mounted into the container |
WIKI_PORT / TASKS_PORT |
8842 / 8851 |
frontend ports |
TZ |
Europe/Moscow |
timezone (matters for cron) |
MIT