Claude-B

Background Claude Code · Telegram bot · REST API · multi-host orchestration.

Run Claude Code as a daemon, talk to it from your phone, and let a four-stage multi-LLM voice pipeline turn rambling voice notes into structured prompts grounded in your live session context.

Install

curl -fsSL https://cb.danimoya.com | bash

Auto-detects npm or docker. Force with --method docker.

Configure

cb init

Interactive wizard — connects the Telegram bot, configures the voice pipeline, writes ~/.claude-b/.env.

The voice pipeline

Most Telegram-AI bots forward your voice note to one model and play the reply back. Claude-B chains four specialised models per round-trip. The differentiator is the middle stage — a prompt optimiser that reads your raw transcript and the last few turns of the target session, then rewrites the request as something Claude Code can actually execute.

Sequence diagram of the Claude-B voice pipeline: Telegram → STT → Optimizer → Claude Code → TTS → Telegram — Voice in, structured prompt out, voice summary back — four LLMs, one trip.

STT — speech to text Voice note arrives from Telegram (.ogg). Transcribed by your chosen STT provider into raw text. Picks up filler words, mishears domain terms, doesn't know what you're working on.
Context gathering Reads the target session's transcript path (cached from the most recent Stop hook) and pulls the last 3 user + 3 assistant turns, plus the working directory and current status. No turn history? The next stage runs ungrounded.
Optimiser — transcript to structured prompt A fast model (default: claude-haiku-4-5) gets the raw transcript and the session context, with a system prompt instructing it to fix STT errors, drop filler, add specificity from context, and preserve intent without inventing requirements. Output: an actionable prompt for Claude Code. You see it on Telegram with confirm / edit / cancel buttons before it runs.
Claude Code — execution The optimised prompt is routed to the right session: a long-running Claude-B session via the daemon, or a live tmux pane via tmux send-keys. Claude Code runs as it would in your terminal — tool calls, edits, the lot.
TTS — text to speech When Claude responds, the result is summarised and synthesised into a voice note that plays back in Telegram. You can listen on your commute instead of staring at a transcript.

Each stage's provider is independent — pick what you have keys for, mix and match, and swap any one without touching the others.

API requirements

The voice pipeline needs three things wired up: an STT provider, a prompt-optimiser provider, and a Telegram bot token. TTS reuses the STT provider's account where supported.

Stage	Provider options	API key env / command	Notes
required Telegram bot	BotFather token	`cb --telegram <TOKEN>`	No bot, no voice in/out. Free; create via @BotFather.
required STT (speech-in)	OpenAI Whisper · Speechmatics · Deepgram	`cb --voice-setup <provider> <KEY>`	Same key handles TTS where the provider supports it (OpenAI, Speechmatics, Deepgram all do).
required Optimiser (LLM)	Anthropic direct · OpenRouter	`cb --ai-provider anthropic <KEY>` `cb --ai-provider openrouter <KEY>`	Default model: `claude-haiku-4-5` (fast, cheap, accurate enough for prompt rewriting). OpenRouter routes to any model.
optional Claude Code execution	Anthropic `claude` CLI	`ANTHROPIC_API_KEY`	Already configured if `claude` works in your terminal. Sonnet / Opus per your subscription.
optional ffmpeg	system package	`apt install ffmpeg`	Required only for TTS (Telegram voice replies). STT-only flows work without it.

Minimum to be voice-active

# 1. Bot
cb --telegram <BOT_TOKEN_FROM_BOTFATHER>

# 2. Speech-to-text + text-to-speech (one provider, one key)
cb --voice-setup openai sk-proj-...

# 3. Prompt optimiser (Anthropic direct — recommended)
cb --ai-provider anthropic sk-ant-...

# Verify
cb --voice-status

What you get

Background Claude Code sessions with async, fire-and-forget prompts
Telegram bot — text and voice in, voice out, inline confirm / edit / cancel
REST API + WebSocket for programmatic access
Live tmux pane integration via send-keys
Multi-host orchestration across machines

Manual install

# npm
npm i -g claude-b

# Docker
docker run -d \
  -v ~/.claude-b:/root/.claude-b \
  --env-file ~/.claude-b/.env \
  ghcr.io/danimoya/claude-b:latest

Pair with a dashboard

Claude-Dashboard is a self-hosted web GUI that surfaces Claude-B's inbox, session transcripts, and live tmux panes — same daemon, browser instead of phone. Both use the recommended single-database tier on HeliosDB-Nano.