Backend Agent

The Backend Agent handles tasks that require tools, file operations, or sustained processing. When the frontend voice LLM determines that a request needs execution, it delegates the goal to the Backend Agent for asynchronous execution; once the result is ready, it naturally returns to the current conversation.

Supported Agents

Backend Agent	Integration Method	Setup Requirements	Skills	Recommendation
None	N/A	Frontend-only mode, no configuration needed	—	★★★★★
Qwen Code	Native ACP	Supports one-click install, requires user configuration	~/.qwen/skills/	★★★★★
OpenCode	Native ACP	Supports one-click install and Bailian configuration	~/.config/opencode/skills/	★★★★★
OpenClaw	Built-in ACP bridge	Supports one-click install and Bailian configuration	~/.openclaw/skills/	★★★★★
Qoder	Native ACP	Supports one-click install, requires user configuration	~/.qoder/skills/	★★★★★
MiniMax Code	Native ACP	Supports one-click install, requires user configuration	Managed by MiniMax Code	★★★★☆
Kimi Code	Native ACP	Supports one-click install, requires user configuration	~/.agents/skills/	★★★★★
Hermes	Native ACP	Supports one-click install, requires user configuration	~/.hermes/skills/	★★★★☆
CodeBuddy	Native ACP	Supports one-click install, requires user configuration	~/.codebuddy/skills/	★★★★☆
Codex	External ACP adapter	Supports one-click install of both the core and adapter, requires user configuration	~/.codex/skills/	★★★★☆
Claude Code	External ACP adapter	Supports one-click install of both the core and adapter, requires user configuration	~/.claude/skills/	★★★★☆
DeepSeek	Native ACP	Supports one-click install, requires a DeepSeek API key	~/.agents/skills/	★★★★☆
Pi	External ACP adapter	Supports one-click install of both the core and adapter, requires user configuration	~/.pi/agent/skills/	★★★★☆
Muse Code	Native MSP adapter	Muse and SDK installed on demand, requires user configuration	Managed by Muse Code	Experimental

Skills install once through qwenaudio skill install (a branded entry point for the standard skills.sh installer) and land automatically in the user-level directories of backends that declare a skills.sh installer above. MiniMax Code and Muse Code manage their own extension systems. See Skill Management.

The recommendation rating reflects the current integration completeness, compatibility, and extent of real-world verification: five stars indicates a fully tested and recommended integration, while four stars indicates ongoing development or incomplete verification of the same scope.

One-Click Install

Uninstalled backend agents can be installed locally with a unified command:

qwenaudio install codex
qwenaudio install deepseek
qwenaudio install minimax
qwenaudio install muse

Before installation, a detection step runs to only fill in missing components: it installs the backend core and, where required, its adapter. Existing components are not reinstalled. Installation does not mean the backend is configured: complete any login and configuration required by the selected backend. In the desktop settings page's "Backend Agent" list, an "Install" button appears at the end of rows for uninstalled backends that support one-click install, using the same installation logic as the CLI.

Existing local Agents and their configuration take priority. Installation only adds missing hosts or adapters, using the official latest tag for npm packages. It does not upgrade or downgrade existing installations. When only the ACP adapter is missing, the desktop button reads “Install adapter”. Minimum compatibility checks remain in place; incompatible installations require a user-managed upgrade.

DeepSeek uses the official CLI's dsh --profile acp entry point (0.1.5 series or newer), without a separate ACP demo. Run dsh web to configure credentials. Without a backend model override, the ACP profile keeps its own model settings; explicit overrides use the standard ACP interface. Gateway-managed independent task delegation and native Session history restoration are not provided yet.

View currently available backend agents:

qwenaudio setup

This command only checks — it does not install, download, or verify credentials. To check only a specific backend or get machine-readable results:

qwenaudio setup --backend codex
qwenaudio setup --json

Choosing a Backend

AGENT_PROTOCOL is an optional configuration. When left empty, the Gateway runs in frontend-only mode, and real-time voice chat remains available; requests requiring backend execution will return a clear explanation without creating a task or guessing results. You can also use qwenaudio --backend none on the command line to explicitly start in frontend-only mode.

AGENT_PROTOCOL=openclaw

OpenCode and OpenClaw support automatic download and installation; after configuring DASHSCOPE_API_KEY and QWEN_AUDIO_AGENT_BACKEND_MODEL, they can automatically connect to Bailian models. Other backends require prior installation and native configuration; qwen-audio-agent will reuse their user-level models, tools, MCPs, Skills, and authentication where the selected protocol exposes them. MiniMax Code and Muse Code keep their own model, provider, authentication, and extension configuration under their control.

To use other agents that support ACP stdio:

AGENT_PROTOCOL=acp
ACP_COMMAND=your-agent
ACP_ARGS=["--acp"]

The command, arguments, display name, and working directory can be configured via ACP_COMMAND, ACP_ARGS, ACP_LABEL, and ACP_WORKSPACE respectively. The generic ACP entry does not provide one-click install; please install it yourself.

Permission Modes

QWEN_AUDIO_AGENT_BACKEND_PERMISSION_MODE can be set to:

• native (default): Permissions are determined and prompted by the backend agent itself; the Gateway only forwards requests as-is.
• full: Grants the highest permissions at startup, allowing the backend to directly execute commands, read and write files without per-action confirmation.

full currently supports OpenCode, Qoder, Qwen Code, MiniMax Code, Kimi Code, Hermes, CodeBuddy, Codex, Claude Code, DeepSeek, and Muse Code; the Gateway will automatically approve permission requests from these backends. OpenClaw's execution authorization is constrained by exec approvals, elevated, and other configuration settings, and cannot be expressed via a single toggle — when full is selected, the Gateway will explicitly refuse to start. The highest permissions amplify the risk of accidental operations and should only be enabled in trusted projects.

Pi is a special case: it has no built-in sandbox or permission approval mechanism, and its adapter pi-acp does not implement ACP session/request_permission. Pi therefore always runs with the equivalent of full permissions regardless of the configured mode — there is no approval step at all, and no permission confirmation appears in the voice session. Use it only in trusted projects and trusted prompt environments.

The current community adapter does not wire ACP mcpServers into Pi, so Gateway Session tools and independent third-layer delegation are unavailable for this backend. Pi handles work in the current Session with its own tools.

Backend Service

To keep the Gateway running, see Run the Gateway. See common settings for configuration and permissions, and backend-specific settings for models, commands, and directories.