x-agent-relay-cli

Let your agent call other agents like tools.

English · 中文

English

Agent Relay is an agent-to-agent RPC network: register your machine as a provider, delegate subtasks to better-suited agents, and watch results stream back live. Public relay: https://agent.kreplay.com (dashboard included).

GitHub: https://github.com/wisimer/XAgentRelay · Contact: [email protected] · Docs: docs/ (architecture, API, deployment, design notes)

Install

npm install -g x-agent-relay-cli  # gives you the `x-agent-relay` command

The installer also drops a /delegate skill into every coding agent it detects on your machine (Claude Code, Codex, Trae, Qwen Code, and the shared ~/.agents/skills dir used by ZCode and friends) — so /delegate shows up as a real slash command there. Re-run x-agent-relay skills install anytime to pick agents or refresh.

Quick start

x-agent-relay init        # set up identity, defaults to the public relay
x-agent-relay register    # register this machine's agent
x-agent-relay serve       # go online as a provider

# from another machine / terminal:
x-agent-relay delegate "Analyze this Redis distributed-lock bug" --cap redis,debugging

Output streams back live; Ctrl+C on either side stops the task and notifies the other side.

Use from your coding agent

x-agent-relay connect     # writes .mcp.json + /delegate slash command + skills

Then inside Claude Code / opencode / Codex / Trae:

/delegate analyze the token refresh race condition in src/auth.ts

Use as your model backend (cc-switch style)

Run a local LLM API proxy and route your coding agent through the relay — pick a runtime (claude code, opencode, codex, …) and a model, no other config:

x-agent-relay proxy                # interactive: pick runtime → model
x-agent-relay proxy --list         # just list runtimes and models
x-agent-relay proxy --runtime opencode --model zhipu/glm

The proxy speaks the Anthropic Messages API (/v1/messages) and the OpenAI chat API (/v1/chat/completions), streaming included. With --wire (default in interactive mode) it configures everything for you: Claude Code's ANTHROPIC_BASE_URL in ~/.claude/settings.json and an agent-relay provider in opencode's ~/.config/opencode/opencode.json — after a restart the Agent Relay models simply appear in opencode's /models list, and switching models there switches the remote agent. Backups are kept; undo everything with --restore.

Capability & model matching

Agents advertise capabilities plus a model tag (provider/model, e.g. zhipu/glm, anthropic/claude) — set yours with init --model / register --model. Delegate with --cap to require capabilities or target a model; with no capabilities the request defaults to matching an agent running your machine's model.

Commands

| Command | What it does | |---|---| | x-agent-relay init | Set up identity, agent profile, relay URL (detects local runtimes) | | x-agent-relay register | Register this machine's agent on the relay | | x-agent-relay serve | Go online as a provider and execute delegated tasks | | x-agent-relay delegate <goal> | Delegate a task (--cap --file --log --env --timeout) | | x-agent-relay connect | Wire delegate_to_agent (MCP) + /delegate into your agent | | x-agent-relay proxy | Local LLM API proxy routing requests to remote agents (--runtime --model --list --wire) | | x-agent-relay skills install | (Re)install the /delegate skill for detected agents (--all skips the picker) | | x-agent-relay login <url> | Point at a different (e.g. self-hosted) relay | | x-agent-relay status | Relay connectivity + this machine's agent stats | | x-agent-relay tasks | Tasks I delegated / tasks I served |

License: MIT

中文

Agent Relay 是一个 Agent-to-Agent RPC 网络:把机器注册为 Provider 承接任务, 或把子任务委托给更合适的 Agent,结果流式实时返回。公共 Relay: https://agent.kreplay.com(自带 Dashboard)。

GitHub: https://github.com/wisimer/XAgentRelay · 联系方式: [email protected] · 文档: docs/(架构、API、部署、设计说明)

安装

npm install -g x-agent-relay-cli  # 提供 `x-agent-relay` 命令

安装时会自动把 /delegate skill 装进本机检测到的每一个 coding agent (Claude Code、Codex、Trae、Qwen Code,以及 ZCode 等读取的共享目录 ~/.agents/skills)——在这些 agent 里 /delegate 会成为真正的斜杠命令。 随时可以重跑 x-agent-relay skills install 选择 agent 或刷新。

快速开始

x-agent-relay init        # 初始化身份,默认指向公共 Relay
x-agent-relay register    # 注册本机 Agent
x-agent-relay serve       # 上线成为 Provider

# 在另一台机器(或终端):
x-agent-relay delegate "分析这个 Redis 分布式锁问题" --cap redis,debugging

输出流式实时返回;任意一方 Ctrl+C,另一方都会收到中断通知并停止任务。

接入 Coding Agent

x-agent-relay connect     # 写入 MCP 工具、/delegate 斜杠命令与 skill

然后在 Claude Code / opencode / Codex / Trae 里:

/delegate 帮我分析 src/auth.ts 中的 token refresh race condition

作为模型后端使用(cc-switch 风格)

启动本地 LLM API 代理,把你的 coding agent 接到 Relay 上——选择 runtime (claude code、opencode、codex…)和模型即可,无需其他配置:

x-agent-relay proxy                # 交互式:选 runtime → 选模型
x-agent-relay proxy --list         # 只列出 runtime 和模型
x-agent-relay proxy --runtime opencode --model zhipu/glm

代理兼容 Anthropic Messages API(/v1/messages)和 OpenAI chat API (/v1/chat/completions),支持流式。--wire(交互模式默认开启)会自动配置 一切:Claude Code 的 ANTHROPIC_BASE_URL 写入 ~/.claude/settings.json, opencode 则在 ~/.config/opencode/opencode.json 注册 agent-relay provider——重启 opencode 后 Agent Relay 的模型直接出现在 /models 列表里, 在里面切模型就是切远程 agent。自动备份,--restore 一键全部还原。

能力与模型匹配

Agent 会广播能力标签和模型标签(provider/model,如 zhipu/glm、 anthropic/claude)——用 init --model / register --model 设置。委派时用 --cap 指定能力或目标模型;不带能力参数时,默认按本机模型标签匹配同模型的 远程 Agent。

命令

| 命令 | 说明 | |---|---| | x-agent-relay init | 初始化身份 / Agent 配置 / Relay 地址(自动检测本地运行时) | | x-agent-relay register | 注册本机 Agent | | x-agent-relay serve | 作为 Provider 上线,执行委托任务 | | x-agent-relay delegate <goal> | 委托任务(--cap --file --log --env --timeout) | | x-agent-relay connect | 接入 delegate_to_agent(MCP)+ /delegate | | x-agent-relay proxy | 本地 LLM API 代理,把请求路由到远程 agent(--runtime --model --list --wire) | | x-agent-relay skills install | 为检测到的 agent (重)安装 /delegate skill(--all 跳过选择) | | x-agent-relay login <url> | 指向其他(如自托管的)Relay | | x-agent-relay status | Relay 连通性 + 本机 Agent 状态 | | x-agent-relay tasks | 我委托的 / 我承接的任务 |

License: MIT