Multi-tenant AI agent server that bridges messaging platforms with AI backends through a fully decoupled architecture.

Each user gets their own isolated AI session. No API keys required — use your existing AI subscription (Claude Max, etc.) and share it across users on any supported messaging platform.

Channel (messaging)     Agent (AI backend)      Runtime (process container)
  ├── WeChat ✓            ├── Claude Code ✓        ├── tmux ✓
  ├── Telegram            ├── Codex                ├── screen
  └── Discord             └── Gemini               └── container

                    ↕                ↕
               [Server orchestration]
            per-user session management

All three dimensions are fully decoupled via Go interfaces. Adding a new channel, agent, or runtime requires zero changes to existing code.

• New Agent (e.g., Codex): implement Agent interface — different agents may use different communication methods (MCP, JSONL polling, API calls)
• New Channel (e.g., Telegram): implement Channel + Presenter — different channels may offer inline keyboards, buttons, or text-only interaction
• New Runtime (e.g., containers): implement Runtime — different runtimes may use Docker, screen, or direct process management

• Claude Code CLI installed and authenticated
• tmux

curl -sSfL https://raw.githubusercontent.com/CMGS/gua/refs/heads/master/scripts/install.sh | bash

This will:

• Download and install gua-server + gua-bridge to /usr/bin/
• Create a systemd service (gua.service)
• Set up work directory at ~/.gua/workspace/

For a specific version: GUA_VERSION=0.2 curl ... | bash

Download from releases and extract manually:

# Apple Silicon
curl -sSfL https://github.com/CMGS/gua/releases/download/v0.1/gua_0.1_Darwin_arm64.tar.gz | tar xz
sudo mv gua-server gua-bridge /usr/local/bin/

Requires Go 1.24+.

make build

# Check version
./bin/gua-server --version

./bin/gua-server setup --backend wechat
# Scan the QR code with WeChat
# Run again for additional accounts

# If installed via install.sh:
sudo systemctl start gua
sudo systemctl enable gua  # auto-start on boot

# Or run directly:
gua-server start \
  --backend wechat \
  --agent claude \
  --work-dir ~/.gua/workspace \
  --bridge-bin /usr/bin/gua-bridge \
  --model sonnet \
  --prompt ./my-custom-rules.md  # optional: appended to init prompt

# List registered bot accounts (~/.gua/<backend>/accounts/)
./bin/gua-server accounts list
./bin/gua-server accounts list --backend wechat

# Remove a bot account
./bin/gua-server accounts remove <account-id>

# List user chat sessions (<work-dir>/sessions/)
./bin/gua-server sessions list --work-dir /path/to/workspace

# Remove a user's session data
./bin/gua-server sessions remove <user-id> --work-dir /path/to/workspace

Global (handled by server, all channels):

Agent CLI passthrough (Claude Code specific, forwarded to terminal):

Channel control (WeChat specific, parsed by Presenter):

/no and /cancel both map to the same action (deny); the distinction is UX — prompts show /no for permission approval and /cancel for TUI menu navigation. Voice commands only take effect when there's an active prompt; otherwise they're sent as normal messages.

gua/
├── version/            Build version info (injected via ldflags)
├── types/              Shared types (MediaFile, Action)
├── utils/              Shared utilities (JSON, MIME, SyncValue, SyncQueue, TempFile)
├── config/             Base prompt (security rules)
├── runtime/            Process container interface
│   └── tmux/           tmux implementation
├── agent/              Agent interface
│   └── claude/         Claude Code implementation
│       ├── protocol/   Bridge socket protocol (envelope types)
│       ├── mcpserver/  Lightweight MCP JSON-RPC server
│       └── bridge/     Bridge binary (MCP server inside Claude Code)
├── channel/            Channel + Presenter interfaces
│   └── wechat/         WeChat implementation
├── libc/wechat/        WeChat iLink protocol library
├── server/             Server orchestration (Channel ↔ Agent routing)
├── cmd/                CLI entry point (setup, start, accounts, sessions)
├── scripts/            Install scripts (Linux systemd setup)
└── bot/                Reference implementations
    ├── weclaw/         WeChat + Claude (ACP/CLI/HTTP) — Go
    └── ccbot/          Telegram + Claude Code (tmux + JSONL) — Python

Interface-driven Architecture: Server orchestration knows nothing about specific agents, channels, or runtimes. The Agent interface defines how to send messages and handle control actions. The Channel interface defines how to receive and deliver messages. The Runtime interface defines how to manage processes. Each implementation decides its own communication method, session model, and capabilities.

Agent CLI Command Whitelist: Each agent declares which CLI commands it supports via CLICommands(). The server forwards matching /commands directly to the agent's terminal; unrecognized commands are sent as normal messages through the agent's standard message protocol.

Presenter Pattern: Channel-specific rendering is encapsulated in a Presenter interface. The same agent response can be rendered as plain text with /select N commands (WeChat), inline keyboard buttons (Telegram), or rich embeds (Discord). The agent layer produces structured data; the presenter decides how to display it.

Four-layer Prompt: Init prompt = config/base.md (security rules) + agent prompt (agent-specific behavior) + channel prompt (channel-specific rules) + presenter.MediaInstructions() (media handling) + optional user prompt (--prompt flag). Each layer is independently managed; user prompt has highest priority (appended last).

The Claude Code agent uses these specific mechanisms (other agents may differ):

• MCP Channel Protocol: Persistent bidirectional communication via Unix socket bridge
• Hook-based Permissions: PermissionRequest and Elicitation hooks intercept prompts before terminal display
• TUI Menu Capture: capture-pane with separator + Esc to boundary detection for CLI command menus
• Watch: FIFO-based pipe-pane streaming for runtime interactive prompt detection

• Tencent WeChat OpenClaw — WeChat iLink Bot API
• weclaw — Pioneer WeChat AI agent bridge, inspired the iLink protocol implementation
• ccbot — Telegram Claude Code bot, demonstrated tmux-based TUI handling and JSONL output monitoring

MIT