Merlion

Merlion is a CLI coding agent that helps you get real work done. You can drive it from your terminal or from WeChat, and point it at basically any OpenAI-compatible model you want.

Install from npm (Recommended)

npm install -g merlion
merlion

Project-local install:

npm install merlion
npx merlion

On first run, Merlion opens a setup wizard for provider/key/model.

Why Merlion

• Works from terminal or WeChat
• Works with OpenAI-compatible providers and custom model endpoints
• Stays simple to set up and use

WeChat Mode

Use your WeChat as the agent inbox.

# first time or token refresh
merlion wechat --login

# daily use
merlion wechat

Inside REPL, you can also trigger QR login directly:

:wechat
/wechat

://wechat in REPL now performs login and immediately enters listening mode (Ctrl+C returns to REPL).

Credentials are stored at ~/.config/merlion/wechat.json. WeChat chat receives final turn replies (plus concise error hints), not internal tool logs. By default WeChat sends only final replies (plus concise error hints). If you want turn-by-turn progress, set MERLION_WECHAT_PROGRESS=1. Set MERLION_WECHAT_PROGRESS_VERBOSE=1 to include extra per-turn tool-batch summaries. Progress updates are capped per request (MERLION_WECHAT_MAX_PROGRESS_UPDATES, default 10) and auto-silenced on server throttling. If a complex task hits turn budget, increase MERLION_WECHAT_MAX_TURNS (default 50). Interactive terminal approvals are not available in WeChat mode; default falls back to --auto-allow. Use --auto-deny if you want risky tools to be blocked.

Common Usage

# one-shot
merlion "read src/index.ts and summarize architecture"

# interactive REPL
merlion

# continue previous session
merlion --resume <session-id>

AGENTS/MERLION Map Automation

Merlion supports layered MERLION.md/AGENTS.md guidance and auto-maintained AUTO sections.

npm run hooks:install
npm run agents:update:staged
npm run agents:lint

Bench

Merlion ships with two benchmark lanes:

• npm run bench:phase0
• npm run bench:bugsinpy

bench:phase0 is the fast fixture-based regression lane. bench:bugsinpy is the separate medium-weight Python bug-fix lane and requires a local BugsInPy clone via MERLION_BUGSINPY_HOME or BUGSINPY_HOME.

---

Merlion（中文）

Merlion 是一个能帮你把活往前推进的 coding agent。 你可以在终端里用它，也可以直接接到微信上用；模型上基本不挑，兼容 OpenAI 协议的都能接。

npm 安装（推荐）

npm install -g merlion
merlion

项目内本地安装：

npm install merlion
npx merlion

首次运行会自动进入配置向导（provider / key / model）。

为什么是 Merlion

• 终端里能用，微信里也能用
• 兼容 OpenAI 协议，基本任何模型都能接
• 安装和上手都尽量简单

微信模式

# 首次登录或 token 过期后
merlion wechat --login

# 日常直接连接
merlion wechat

在 REPL 里也可以直接弹二维码登录：

:wechat
/wechat

REPL 里的 :wechat / /wechat 现在会“登录后直接进入监听模式”（按 Ctrl+C 返回 REPL）。

凭据保存位置：~/.config/merlion/wechat.json。 微信端默认只接收每轮最终答复（外加简短错误提示），不会推送内部工具日志。 如需开启逐轮进度推送，可设置 MERLION_WECHAT_PROGRESS=1。 如需更细粒度（含每轮工具批次汇总），可设置 MERLION_WECHAT_PROGRESS_VERBOSE=1。 每次请求的进度推送有上限（MERLION_WECHAT_MAX_PROGRESS_UPDATES，默认 10），遇到服务端限流会自动静默后续进度。 复杂任务若触达轮次上限，可调大 MERLION_WECHAT_MAX_TURNS（默认 50）。 微信模式不支持终端交互式审批；默认会回退到 --auto-allow，如需严格阻止高风险工具请使用 --auto-deny。

常用命令

merlion "read src/index.ts and summarize architecture"
merlion
merlion --resume <session-id>

AGENTS / MERLION 地图自动维护

npm run hooks:install
npm run agents:update:staged
npm run agents:lint

Bench / 评测

Merlion 现在有两条评测链路：

• npm run bench:phase0
• npm run bench:bugsinpy

bench:phase0 是快速的 fixture 型回归评测。 bench:bugsinpy 是单独的中量级 Python 真实缺陷评测，需要通过 MERLION_BUGSINPY_HOME 或 BUGSINPY_HOME 指向本地 BugsInPy 仓库。