Claude Code and Codex CLI as your personal Telegram assistant.
Persistent memory. Scheduled tasks. Live streaming. Docker sandboxing.
Uses only the official CLIs. Nothing spoofed, nothing proxied.

Quick start · Why ductor? · Features · Prerequisites · How it works · Commands · Docs · Contributing

---

ductor runs on your machine, uses your existing Claude or Codex subscription, and remembers who you are between conversations. One Markdown file is the agent's memory. One folder (~/.ductor/) holds everything. pipx install ductor, done.

You can schedule cron jobs, set up webhooks, and let the agent check in on its own with heartbeat prompts. Responses stream live into Telegram. Sessions survive restarts.

_{Left: /start onboarding screen — Right: Quick action buttons generated dynamically by the agent}

Quick start

pipx install ductor
ductor

The setup wizard walks you through the rest.

Why ductor?

I tried a bunch of CLI wrappers and Telegram bots for Claude and Codex. Most were either too complex to set up, too hard to modify, or got people banned because they spoofed headers and forged API requests to impersonate the official CLI.

ductor doesn't do that.

• Spawns the real CLI binary as a subprocess. No token interception, no request forging
• Uses only official rule files: CLAUDE.md and AGENTS.md
• Memory is one Markdown file. No RAG, no vector stores
• One channel (Telegram), one Python package, one command

The agents are good enough now that you can steer them through their own rule files. I don't need a RAG system to store memories -a single Markdown file that tracks what I like, what I don't, and what I'm working on is plenty. I can reach them from Telegram instead of a terminal.

I picked Python because it's easy to modify. The agents can write their own automations, receive webhooks (new email? parse it and ping me), set up scheduled tasks. All controlled from your phone.

Features

Core

• Responses stream in real-time -ductor edits the Telegram message live as text arrives
• Switch between Claude Code and Codex mid-conversation with /model
• Sessions survive bot restarts
• @opus explain this temporarily switches model without changing your default
• Send images, PDFs, voice messages, or videos -ductor routes them to the right tool
• Agents can send [button:Yes] [button:No] inline keyboards back to you
• Persistent memory across sessions, stored in one Markdown file

Automation

• Cron jobs -recurring tasks with cron expressions and timezone support. Each job runs as its own subagent with a dedicated workspace and memory file (plus optional per-job quiet hours and dependency locks)
• Webhooks -HTTP endpoints with Bearer or HMAC auth. Two modes: wake injects a prompt into your active chat, cron_task runs a separate task session. Works with GitHub, Stripe, or anything that sends POST
• Heartbeat -the agent checks in periodically during active sessions. Quiet hours respected
• Cleanup -daily retention cleanup for telegram_files/ and output_to_user/

Example: a cron job

Tell the agent: "Check Hacker News every morning at 8 and send me the top AI stories."

ductor creates a task folder with everything the subagent needs:

~/.ductor/workspace/cron_tasks/hn-ai-digest/
    CLAUDE.md              # Agent rules (managed by ductor)
    AGENTS.md              # Same rules for Codex
    TASK_DESCRIPTION.md    # What the agent should do
    hn-ai-digest_MEMORY.md # The subagent's own memory across runs
    scripts/               # Helper scripts if needed

At 8:00 every morning, ductor starts a fresh session in that folder. The subagent reads the task, does the work, writes what it learned to memory, and posts the result to your chat. It is context-isolated from your main conversation and memory.

Example: a webhook wake call

Your CI fails. A webhook in wake mode injects the payload into your active chat. Your agent sees it with full history and memory and responds.

POST /hooks/ci-failure -> "CI failed on branch main: test_auth.py::test_login timed out"
-> Agent reads this, checks the code, tells you what went wrong

Infrastructure

• ductor service install -Linux systemd user service (start on boot, restart on crash)
• Docker sandbox image (built via Dockerfile.sandbox) -both CLIs have full filesystem access by default, so a container keeps your host safe
• /upgrade checks PyPI, offers in-chat upgrade, then restarts automatically on success
• Supervisor with PID lock. Exit code 42 triggers restart
• Prompt injection detection, path traversal checks, per-user allowlist

Developer experience

• First-run wizard detects your CLIs, walks through config, seeds the workspace
• New config fields merge automatically on upgrade
• /diagnose shows system diagnostics (version/provider/model, Codex cache status, recent logs), /status shows session stats
• /stop terminates the active run and drains queued messages, /new starts a fresh session
• /showfiles lets you browse ~/.ductor/ as a clickable file tree inside Telegram
• Messages sent while the agent is working show [Message in queue...] with a cancel button
• Bundled skills (e.g. skill-creator) are symlinked into the workspace and stay current with the installed version

Prerequisites

Requirement	Details
Python 3.11+	python3 --version
pipx	pip install pipx (recommended) or pip
One CLI installed	Claude Code or Codex CLI
CLI authenticated	claude auth or codex auth
Telegram Bot Token	From @BotFather
Your Telegram User ID	From @userinfobot
Docker (optional)	Recommended for sandboxed execution

Detailed platform guides: Installation (Linux, macOS, WSL, Windows, VPS)

How it works

You (Telegram)
    |
    v
ductor (aiogram)
    |
    ├── AuthMiddleware (user allowlist)
    ├── SequentialMiddleware (per-chat lock + queue tracking)
    |
    v
Orchestrator
    |
    ├── Command Router (/new, /model, /stop, ...)
    ├── Message Flow -> CLIService -> claude / codex subprocess
    ├── CronObserver -> Scheduled task execution
    ├── HeartbeatObserver -> Periodic background checks
    ├── WebhookObserver -> HTTP endpoint server
    ├── CleanupObserver -> Daily file retention cleanup
    └── UpdateObserver -> PyPI version check
    |
    v
Streamed response -> Live-edited Telegram message

ductor spawns the CLI as a child process and parses its streaming output. The Telegram message gets edited live as text arrives. Sessions are stored as JSON. Background systems run as asyncio tasks in the same process.

Workspace

Everything lives in ~/.ductor/.

~/.ductor/
    config/config.json           # Bot config (token, user IDs, model, Docker, timezone)
    sessions.json                # Active sessions per chat
    cron_jobs.json               # Scheduled task definitions
    webhooks.json                # Webhook endpoint definitions
    CLAUDE.md                    # Agent rules (auto-synced)
    AGENTS.md                    # Same rules for Codex (auto-synced)
    logs/agent.log               # Rotating log file
    workspace/
        memory_system/
            MAINMEMORY.md        # The agent's long-term memory about you
        cron_tasks/              # One subfolder per scheduled job
        skills/
            skill-creator/       # Bundled skill (symlinked from package)
        tools/
            cron_tools/          # Add, edit, remove, list cron jobs
            webhook_tools/       # Add, edit, remove, test webhooks
            telegram_tools/      # Process files, transcribe audio, read PDFs
            user_tools/          # Custom scripts the agent builds for you
        telegram_files/          # Downloaded media, organized by date
        output_to_user/          # Files the agent sends back to you

Plain text, JSON, and Markdown. No databases, no binary formats.

Configuration

Config lives in ~/.ductor/config/config.json. The wizard creates it on first run:

ductor  # wizard creates config interactively

Key fields: telegram_token, allowed_user_ids, provider (claude or codex), model, docker.enabled, user_timezone, cleanup. Full schema in docs/config.md.

CLI Parameters

Configure provider-specific CLI parameters in config.json:

{
  "cli_parameters": {
    "claude": [],
    "codex": ["--chrome"]
  }
}

Parameters are appended to CLI commands for the respective provider.

Advanced Cron Task Configuration

Cron tasks support per-task execution overrides:

{
  "provider": "codex",
  "model": "gpt-5.2-codex",
  "reasoning_effort": "high",
  "cli_parameters": ["--chrome"],
  "quiet_start": 22,
  "quiet_end": 7,
  "dependency": "nightly-reports"
}

All fields are optional and fall back to global config values if not specified.

Commands

Command	Description
/new	Start a fresh session
/stop	Stop active agent execution and discard queued messages
/model	Switch AI model (interactive keyboard)
/model opus	Switch directly to a specific model
/status	Session info, tokens, cost, auth status
/memory	View persistent memory
/cron	View/manage scheduled tasks (toggle enable/disable)
/showfiles	Browse ~/.ductor/ as an interactive file tree
/info	Project links and version info
/upgrade	Check for updates and show upgrade prompt
/restart	Restart the bot
/diagnose	Show system diagnostics and recent logs
/help	Command reference

Documentation

Document	Description
Installation	Platform-specific setup (Linux, macOS, WSL, Windows, VPS)
Developer Quickstart	Fast onboarding for contributors and junior devs
Automation	Cron jobs, webhooks, heartbeat
Configuration	Full config schema and options
Architecture	System design and runtime flow
Module reference	Per-subsystem documentation

Disclaimer

ductor runs the official CLI binaries from Anthropic and OpenAI. It does not modify API calls, spoof headers, forge tokens, or impersonate clients. Every request comes from the real CLI process.

Terms of Service can change. Automating CLI interactions may be a gray area depending on how providers interpret their rules. We built ductor to follow intended usage patterns, but can't guarantee it won't lead to account restrictions.

Use at your own risk. Check the current ToS before deploying:

• Anthropic Terms of Service
• OpenAI Terms of Use

Not affiliated with Anthropic or OpenAI.

Contributing

git clone https://github.com/PleasePrompto/ductor.git
cd ductor
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
ruff check .
mypy ductor_bot

Zero warnings, zero errors. See CLAUDE.md for conventions.

License

MIT