ductor

If you want to control Claude Code, Google's Gemini CLI, or OpenAI's Codex CLI via Telegram or Matrix, build automations, or manage multiple agents easily — ductor is the right tool for you. The messaging layer is modular: Telegram and Matrix ship today, and new transports plug into the same transport-agnostic core.

ductor runs on your machine and sends simple console commands as if you were typing them yourself, so you can use your active subscriptions (Claude Max, etc.) directly. No API proxying, no SDK patching, no spoofed headers. Just the official CLIs, executed as subprocesses, with all state kept in plain JSON and Markdown under ~/.ductor/.

Quick start

pipx install ductor    # or: uv tool install ductor
ductor

The onboarding wizard handles CLI checks, transport setup (Telegram or Matrix), timezone, optional Docker, and optional background service install.

Requirements: Python 3.11+, at least one CLI installed (claude, codex, gemini, or agy), and either:

• a Telegram Bot Token from @BotFather, or
• a Matrix account on a homeserver (homeserver URL, user ID, password/access token)

For Matrix support: ductor install matrix — see Matrix setup guide.

Detailed setup: docs/installation.md

How chats work

ductor gives you multiple ways to interact with your coding agents. Each level builds on the previous one.

1. Single chat (your main agent)

This is where everyone starts. You get a private 1:1 chat with your bot (Telegram or Matrix). Every message goes to the CLI you have active (claude, codex, gemini, or agy), responses stream back in real time.

You:   "Explain the auth flow in this codebase"
Bot:   [streams response from Claude Code]

You:   /model
Bot:   [interactive model/provider picker]

You:   "Now refactor the parser"
Bot:   [streams response, same session context]

This single chat is all you need. Everything else below is optional.

2. Groups with topics (multiple isolated chats)

Telegram: Create a group, enable topics (forum mode), and add your bot. Matrix: Invite the bot to multiple rooms — each room is its own context.

Every topic (Telegram) or room (Matrix) becomes an isolated chat with its own CLI context.

Group: "My Projects"
  ├── General           ← own context (isolated from your single chat)
  ├── Topic: Auth       ← own context
  ├── Topic: Frontend   ← own context
  ├── Topic: Database   ← own context
  └── Topic: Refactor   ← own context

That's 5 independent conversations from a single group. Your private single chat stays separate too — 6 total contexts, all running in parallel.

Each topic can use a different model. Run /model inside a topic to change just that topic's provider.

All chats share the same ~/.ductor/ workspace — same tools, same memory, same files. The only thing isolated is the conversation context.

Telegram note: The Bot API has no method to list existing forum topics. ductor learns topic names from forum_topic_created and forum_topic_edited events — pre-existing topics show as "Topic #N" until renamed. This is a Telegram limitation, not a ductor limitation.

3. Named sessions (extra contexts within any chat)

Need to work on something unrelated without losing your current context? Start a named session. It runs inside the same chat but has its own CLI conversation.

You:   "Let's work on authentication"        ← main context builds up
Bot:   [responds about auth]

/session Fix the broken CSV export            ← starts session "firmowl"
Bot:   [works on CSV in separate context]

You:   "Back to auth — add rate limiting"     ← main context is still clean
Bot:   [remembers exactly where you left off]

@firmowl Also add error handling              ← follow-up to the session

Sessions work everywhere — in your single chat, in group topics, in sub-agent chats. Think of them as opening a second terminal window next to your current one.

4. Background tasks (async delegation)

Any chat can delegate long-running work to a background task. You keep chatting while the task runs autonomously. When it finishes, the result flows back into your conversation.

You:   "Research the top 5 competitors and write a summary"
Bot:   → delegates to background task, you keep chatting
Bot:   → task finishes, result appears in your chat

You:   "Delegate this: generate reports for all Q4 metrics"
Bot:   → explicitly delegated, runs in background
Bot:   → task has a question? It asks the agent → agent asks you → you answer → task continues

Each task gets its own memory file (TASKMEMORY.md) and can be resumed with follow-ups.

5. Sub-agents (fully isolated second agent)

Sub-agents are completely separate bots — own chat, own workspace, own memory, own CLI auth, own config settings (heartbeat, timeouts, model defaults, etc.). Each sub-agent can use a different transport (e.g. main on Telegram, sub-agent on Matrix).

ductor agents add codex-agent    # creates a new bot (needs its own BotFather token)

Your main chat (Claude):        "Explain the auth flow"
codex-agent chat (Codex):       "Refactor the parser module"

Sub-agents live under ~/.ductor/agents/<name>/ with their own workspace, tools, and memory — fully isolated from the main agent.

You can delegate tasks between agents:

Main chat:  "Ask codex-agent to write tests for the API"
  → Claude sends the task to Codex
  → Codex works in its own workspace
  → Result flows back to your main chat

Comparison

How it all fits together

~/.ductor/                          ← shared workspace (tools, memory, files)
  │
  ├── Single chat                   ← main agent, private 1:1
  │     ├── main context
  │     └── named sessions
  │
  ├── Group: "My Projects"          ← same agent, same workspace
  │     ├── General (own context)
  │     ├── Topic: Auth (own context, own model)
  │     ├── Topic: Frontend (own context)
  │     └── each topic can have named sessions too
  │
  └── agents/codex-agent/           ← sub-agent, fully isolated workspace
        ├── own single chat
        ├── own group support
        ├── own named sessions
        └── own background tasks

Features

• Multi-transport — run Telegram and Matrix simultaneously, or pick one
• Multi-language — UI in English, Deutsch, Nederlands, Français, Русский, Español, Português
• Real-time streaming — live message edits (Telegram) or segment-based output (Matrix)
• Telegram reasoning + tool UX controls — optional reasoning stream, live tool progress, and separate thinking indicator controls
• Provider switching — /model to change provider/model (never blocks, even during active processes)
• Persistent memory — plain Markdown files that survive across sessions
• Memory maintenance — pre-compaction flush, optional reflection cadence, and LLM-driven compaction
• Cron jobs — in-process scheduler with timezone support, per-job overrides, result routing to originating chat
• Webhooks — wake (inject into active chat) and cron_task (isolated task run) modes
• Heartbeat — proactive checks with per-target settings, group/topic support, chat validation
• Image processing — auto-resize and WebP conversion for incoming images (configurable)
• Media transcription hooks — configurable external audio/video transcription commands for bundled media tools
• Notification routing — startup/upgrade lifecycle messages can target specific chats/topics
• **Task prio