Claudio is an adapter for Claude Code CLI and Telegram. It makes a tunnel between a private network and Telegram's API. So that users can chat with Claude Code remotely, in a safe way.
+---------------------------------------------+
| Remote Machine |
| |
| |
+----------+ | +---------+ +-----------------+ |
| Telegram |<--------+--->| Claudio |<----->| Claude Code CLI | |
+----------+ | +---------+ +-----------------+ |
+---------------------------------------------+
Claudio starts a local HTTP server that listens on port 8421, and creates a tunnel using cloudflared. When the user sends a message from Telegram, it's sent to <cloudflare-tunnel-url>/telegram/webhook and forwarded to the Claude Code CLI.
The user message is passed as a one-shot prompt, along with some context to maintain continuity. All messages are kept in the database, with the last 20 used as conversation context (configurable via MAX_HISTORY_LINES).
After Claude Code finishes, it outputs a JSON response to stdout. Claudio parses the result text and token usage stats, then forwards the response to the Telegram API.
The HTTP server includes several reliability mechanisms:
- Message queue — Incoming webhooks are queued per-chat (max 5) and processed serially, so multiple rapid messages don't spawn overlapping Claude sessions.
- Deduplication — Telegram may retry webhook deliveries; Claudio tracks the last 1,000
update_ids to silently drop duplicates. - Body size limit — Requests larger than 1 MB are rejected to prevent abuse.
- Graceful shutdown — On
SIGTERM, the server stops accepting new requests and waits for active handlers to finish before exiting. - Fallback model — If Claude fails with the selected model, it automatically retries with Haiku as a fallback (unless Haiku is already selected).
The whole purpose of Claudio is to give you remote access to Claude Code from your Telegram app — without breaking Anthropic's Terms of Service. This means Claude Code runs with full permissions by design:
--tools + --allowedTools) AND --disable-slash-commands
Since there's no human in front of the terminal to approve permission prompts, Claude Code must run autonomously. Rather than using --dangerously-skip-permissions, Claudio explicitly lists the tools Claude can use and auto-approves them — excluding interactive-only tools like AskUserQuestion and Chrome. Claudio mitigates the risk through: webhook secret validation (HMAC), single authorized chat_id, and binding the HTTP server to localhost only (external access goes through cloudflared).
- Claude Code CLI (with Pro/Max subscription)
- Linux/macOS/WSL
- Telegram bot token
- Homebrew (macOS only, for installing dependencies)
python3,curlsqlite3,jq,cloudflared(auto-installed if missing)
- Clone the repository and run the install command:
git clone https://github.com/edgarjs/claudio.git
cd claudio
./claudio installThis will:
- Install dependencies (
sqlite3,jq,cloudflared,fastembed) if not already present - Create a symlink at
~/.local/bin/claudioso you can runclaudiofrom anywhere - Set up a Cloudflare named tunnel (permanent URL, requires a free Cloudflare account)
- Install a systemd/launchd service
- Enable loginctl linger on Linux (so the service survives logout on headless systems)
Note: If
~/.local/binis not in your PATH, you'll need to add it. See the installation output for instructions.
- Set up Telegram bot
In Telegram, message @BotFather with /newbot and follow instructions to create a bot. At the end, you'll be given a secret token, copy it and then run:
claudio telegram setupPaste your Telegram token when asked, and press Enter. Then, send a /start message to your bot from the Telegram account that you'll use to communicate with Claude Code.
The setup wizard will confirm when it receives the message and finish. Once done, the service restarts automatically, and you can start chatting with Claude Code.
For security, only the
chat_idcaptured during setup is authorized to send messages.
A cron job runs every 5 minutes to monitor the webhook endpoint. It verifies the webhook is registered and re-registers it if needed. If the server is unreachable, it auto-restarts the service (throttled to once per 3 minutes, max 3 attempts). After exhausting restart attempts without recovery, it sends a Telegram alert and stops retrying until the server responds with HTTP 200. The restart counter auto-clears when the health endpoint returns HTTP 200. You can also reset it manually by deleting
$HOME/.claudio/.last_restart_attemptand$HOME/.claudio/.restart_fail_count.The health check also monitors: orphan
claude/nodeprocesses (kills them after 30 minutes), disk usage (alerts above 90%), log file sizes (rotates files over 10MB), and backup freshness (alerts if the last backup is older than 2 hours). These thresholds are configurable via environment variables.
Check the service and webhook status:
claudio statusTo update Claudio to the latest stable release:
claudio updateOn Linux,
claudio updatealso enables loginctl linger for existing installs upgrading to this version.
To view recent logs (default 50 lines):
claudio logFollow logs in real time with -f, or change the line count with -n:
claudio log -f
claudio log -n 100To stop and remove the service run:
claudio uninstallIf you want to remove the $HOME/.claudio directory completely, run:
claudio uninstall --purgeOn Linux, uninstall disables loginctl linger if no other user services remain enabled.
Claudio uses Haiku by default. If you want to switch to another model, just send the name of the model as a command: /opus, /sonnet, or /haiku. And then continue chatting.
Claudio supports voice messages in both directions using ElevenLabs. Set ELEVENLABS_API_KEY in your service.env to enable voice features (ELEVENLABS_VOICE_ID defaults to Chris if not set). When you send a voice message, Claudio transcribes it (STT), processes it, and responds with a voice message (as a reply). If TTS fails, it falls back to a text-only response. Text messages always get text-only responses. If ELEVENLABS_API_KEY is not set, voice messages will be rejected with an error reply.
You can send photos or image files directly to Claudio. Include an optional caption to tell Claude what to do with the image, or send it without a caption and Claude will describe it.
- Supported formats: JPEG, PNG, GIF, WebP
- Sending as photo: Telegram compresses images automatically
- Sending as document: Attach an image file for lossless quality
- Size limit: 20 MB (Telegram bot API constraint)
Images are validated (magic byte verification, size check) and stored temporarily during processing, then deleted immediately after Claude responds.
You can send documents (PDF, text files, CSV, code files, etc.) to Claudio. Include an optional caption to tell Claude what to do with the file, or send it without a caption and Claude will read and summarize it.
- Any file type supported by Telegram's document upload (files with image MIME types are routed through the image pipeline instead)
- Size limit: 20 MB (Telegram bot API constraint)
- Claude reads the file directly from disk — text-based formats (PDF, CSV, code, plain text) work best; binary files may produce limited results
- Files are stored temporarily during processing, then deleted immediately after Claude responds
Parallel work (reviews, research, etc.) is handled by Claude Code's built-in Task tool (subagents). No custom agent infrastructure is needed — Claude natively spawns subagents, manages their lifecycle, and collects results within a single claude -p invocation.
Claudio can automatically back up its $HOME/.claudio/ directory using rsync-based rotating backups.
# One-off backup
claudio backup /path/to/destination
# Install hourly cron job (default: 24 hourly, 7 daily snapshots)
claudio backup cron install /path/to/destination
# Custom retention
claudio backup cron install /path/to/destination --hours 48 --days 14
# Check backup status
claudio backup status /path/to/destination
# Remove cron job
claudio backup cron uninstallTo customize Claude's behavior, use ~/.claude/CLAUDE.md (Claude Code's built-in configuration file). Instructions there are loaded automatically by Claude Code on every invocation and persist across updates.
Claudio stores its configuration and other files in $HOME/.claudio/. Settings are managed in the service.env file within that directory. After any manual edits, restart to apply your changes:
claudio restartThe following variables can be set in $HOME/.claudio/service.env:
Server
PORT— HTTP server listening port. Default:8421.
Claude
MODEL— Claude model to use. Acceptshaiku,sonnet, oropus. Default:haiku. Can also be changed at runtime via Telegram commands/haiku,/sonnet,/opus.MAX_HISTORY_LINES— Number of recent messages used as conversation context. Default:20.
Telegram
TELEGRAM_BOT_TOKEN— Telegram Bot API token. Set automatically duringclaudio telegram setup.TELEGRAM_CHAT_ID— Authorized Telegram chat ID. Only messages from this chat are processed. Set automatically duringclaudio telegram setup.WEBHOOK_URL— Public URL where Telegram sends webhook updates (e.g.https://claudio.example.com). Set automatically when using a named tunnel.WEBHOOK_SECRET— HMAC secret for validating incoming webhook requests. Auto-generated on first run if not set.WEBHOOK_RETRY_DELAY— Seconds between webhook registration retry attempts. Default:60.
Voice (TTS/STT)
ELEVENLABS_API_KEY— API key for ElevenLabs. Required for voice messages (both TTS and STT).ELEVENLABS_VOICE_ID— ElevenLabs voice ID to use for TTS. Default:iP95p4xoKVk53GoZ742B(Chris).ELEVENLABS_MODEL— ElevenLabs TTS model. Default:eleven_multilingual_v2.ELEVENLABS_STT_MODEL— ElevenLabs STT model. Default:scribe_v1.
Memory
MEMORY_ENABLED— Enable/disable the cognitive memory system. Default:1.MEMORY_EMBEDDING_MODEL— Sentence-transformers model for memory embeddings. Default:sentence-transformers/all-MiniLM-L6-v2.MEMORY_CONSOLIDATION_MODEL— Claude model used for memory consolidation. Default:haiku.
Health Check
DISK_USAGE_THRESHOLD— Disk usage percentage to trigger alerts. Default:90.LOG_MAX_SIZE— Maximum log file size in bytes before rotation. Default:10485760(10 MB).BACKUP_MAX_AGE— Maximum backup age in seconds before alerting. Default:7200(2 hours).BACKUP_DEST— Backup destination path for freshness checks. Default:/mnt/ssd(customize to your backup location).
Tunnel
TUNNEL_NAME— Name of the Cloudflare named tunnel. Set duringclaudio install.TUNNEL_HOSTNAME— Hostname for the named tunnel (e.g.claudio.example.com). Set duringclaudio install.
Most of these variables are configured automatically by
claudio installandclaudio telegram setup. Manual editing is only needed for fine-tuning.
Claudio uses BATS (Bash Automated Testing System) for integration tests.
# Run all tests
bats tests/
# Run a specific test file
bats tests/db.batsCompleted
- Webhook signature validation
- Race conditions — Migrated from JSONL to SQLite
- Basic integration tests
- Input validation for model commands
- ShellCheck linting for scripts
- Cloudflared URL detection fails silently after 30 seconds
- Webhook health check cron job
- No retries with backoff for Telegram API
- Health check (
/health) doesn't verify actual system state - Auto-install dependencies (
sqlite3,jq,cloudflared) duringclaudio install - Show webhook registration failure reason instead of generic warning
- Add
claudioto$PATHvia symlink during install - Environment variables documentation
- Image uploads
- Voice messages from bot (TTS)
- Voice messages from human (STT)
- File uploads
- Parallel work via Claude Code's built-in Task tool (subagents)
- Cognitive memory system (ACT-R activation scoring, embedding-based retrieval)
- Automated backup system (hourly/daily rotating backups with rsync)
Future
- Support for editing messages and reactions
- Rate limiting
- Support for group chats
Claudio is licensed under the Apache License 2.0. See the LICENSE file for details.
Contributions are welcome! Please read the CONTRIBUTING file for details on how to contribute.
Claudio is built on top of Claude and Claude Code.
