Codecast

One bot to manage AI coding agents across all your machines

---

Send a message from Discord, Telegram, or Lark. It reaches an AI coding agent on your GPU server, cloud VM, or any SSH-accessible machine — and streams the response back in real time.

How is this different from the official Claude Discord bot?

The official Claude bot gives you a single conversation with Claude in the cloud. Codecast is fundamentally different:

	Official Claude Bot	Codecast
Where Claude runs	Anthropic's cloud	Your own machines (GPU servers, VMs, dev boxes)
Session management	One conversation	Multiple persistent sessions across multiple machines
Code access	No file system access	Full access to your projects via Claude CLI
AI backends	Claude only	Claude, Codex (OpenAI), Gemini CLI, OpenCode
Chat platforms	Discord only	Discord, Telegram, Lark (Feishu)
Multiplexing	One bot = one conversation	One bot = many machines x many sessions

Codecast turns your chat app into a remote control panel for AI coding agents running on your infrastructure.

Architecture

You (Discord / Telegram / Lark)
 |
 v
+------------------+          SSH tunnel          +------------------+
|    Head Node     |  ----------------------->   |     Daemon       |
|    (Python)      |         JSON-RPC + SSE       |     (Rust)       |
|                  |                              |                  |
|  - Bot adapters  |    Can manage multiple       |  - Axum HTTP     |
|  - Engine        |    machines in parallel       |  - Session pool  |
|  - SSH tunnels   |                              |  - CLI adapters  |
+------------------+                              +--------+---------+
                                                           |
                                                    stdin/stdout
                                                           |
                                                  +--------v---------+
                                                  |   AI CLI Agent   |
                                                  |  Claude / Codex  |
                                                  | Gemini / OpenCode|
                                                  +------------------+

A single Head Node connects to any number of remote machines. Each machine runs a lightweight Rust daemon that manages AI CLI processes. Sessions persist across disconnects — detach from your phone, resume from your laptop.

Quick Start

pip install codecast
cp config.example.yaml ~/.codecast/config.yaml
$EDITOR ~/.codecast/config.yaml   # add machines + bot token
codecast

Then open your chat app and type /start my-server ~/projects/myapp.

Key Features

Multi-Machine Session Management

Connect any number of servers. Each session has a human-friendly name (e.g., bright-falcon). Detach with /exit, resume with /resume bright-falcon from any device.

Multiple AI Backends

Not just Claude. Start sessions with different AI CLIs:

/start my-server ~/project --cli codex
/start gpu-box ~/ml-project --cli gemini
/start dev-vm ~/webapp --cli claude

Three Chat Platforms

• Discord — slash commands, autocomplete, interactive buttons
• Telegram — inline keyboards, file sharing, HTML formatting
• Lark (Feishu) — rich text messages, WebSocket connection

Interactive Questions

When an AI agent asks you a question with options, Codecast presents it with native UI elements — buttons on Discord, inline keyboards on Telegram — so you can tap to respond instead of typing.

Permission Modes

Control how much autonomy the AI agent has:

Mode	Behavior
auto	Full autonomy — no confirmations
code	Auto-accept file edits, confirm shell commands
plan	Read-only analysis
ask	Confirm every action

Real-Time Streaming

Responses stream back as the AI types. Choose how tool calls are displayed:

• timer (default) — clean "Working 5s" indicator, results sent at end
• append — see each tool call as it happens
• batch — summary of all tool calls at end

SSH Security

The daemon binds to 127.0.0.1 only — never exposed to the internet. All communication goes through SSH tunnels. The daemon binary is auto-deployed to remote machines via SCP.

Commands

Command	What it does
/start <machine> <path>	Start a new session (add --cli codex etc.)
/resume <name>	Resume a detached session
/new	New session, same directory
/exit	Detach (process keeps running)
/stop	Interrupt current operation
/model <name>	Switch AI model
/mode <auto|code|plan|ask>	Switch permission mode
/tool-display <timer|append|batch>	Switch tool display mode
/ls machine	List machines
/ls session	List sessions
/status	Current session info
/health	Daemon health check

See the full command reference for all commands.

Documentation

Full documentation: chivier.github.io/codecast (English & Chinese)

Guide	Description
Getting Started	Installation, first session walkthrough
Configuration	Config format, peers, bot tokens
Commands	Every command with examples
Architecture	System design and internals

Requirements

• Python 3.11+ — head node
• SSH access — to remote machine(s) with an AI CLI installed (Claude, Codex, Gemini, or OpenCode)
• Bot token — for Discord, Telegram, and/or Lark

The Rust daemon binary is auto-deployed. To build from source:

cargo build --release
# Binary: target/release/codecast-daemon

License

MIT