laohuang

A minimal coding-agent CLI for local software work.

Install

Recommended standalone install (bundled Node.js; Bash required):

curl -fsSL https://github.com/hxr223/laoHuangCode/releases/latest/download/install.sh | sh

On Windows, install Git for Windows first, then run in PowerShell:

irm https://github.com/hxr223/laoHuangCode/releases/latest/download/install.ps1 | iex

The installer verifies the download and startup, then configures the user PATH. Open a new terminal or apply the PATH command printed by the installer. Standalone packages target macOS, glibc Linux and Windows on x64 and ARM64. See installation details.

Alternatively, install with npm (Node.js >=22.19.0 and Bash required; npm's global executable directory must already be on PATH):

npm install -g laohuang

Usage

laohuang --help
laohuang --version
laohuang config set --provider deepseek --model deepseek-v4-flash
laohuang doctor
laohuang update

Run laohuang in a project directory to start an interactive session. The CLI loads local project instructions, uses saved provider profiles, and keeps file and shell tool execution rooted in the current project.

Model providers come from the installed pi-ai API-key catalog. In pi-ai 0.83.0 the product exposes 35 available API-key providers, excluding Amazon Bedrock, Google Vertex, OAuth-only providers, and OpenAI Codex. Use /providers to see available, configured, and verified status independently.

/login <provider> stores API-key credentials, /logout <provider> removes them, and /model opens a searchable list of all available models from configured providers, including credentials supplied through environment variables. Search by provider ID, model ID, or model name; scroll to browse the full list. If no models are available, the CLI points to /login. /model <provider> narrows the list, /model <provider> <model> switches directly, and /apikey remains a compatibility alias. Switching models affects only the current session, not the default profile.

For standalone installations, laohuang update reruns the bundled installer against the latest GitHub release, verifies it before switching the launcher, and retains the previous version. Restart laohuang after updating.

For npm installations, laohuang update updates the running installation to the latest stable release and verifies its installed version. Restart the CLI after updating. It supports global npm installations and direct project or workspace npm dependencies; local updates change the owning manifest and lockfile while preserving the dependency bucket and ordinary version-prefix style. It does not downgrade or update a different installation. Source checkouts, linked packages, temporary npx installations, and other package managers require their respective update methods. Installation failures are reported without claiming that npm rolled back changes.

/name MCP design saves the current session name, and /name displays it. Names survive resume and appear in session lists. Clones inherit the current name; forks inherit the name at their history boundary.

/copy copies the latest recorded assistant response's text with its Markdown intact, excluding reasoning and tool results. During streaming it copies the previous recorded response. Local sessions use the operating system clipboard; SSH sessions send an OSC 52 request to the user's terminal. A sent request is not confirmation that the terminal accepted it. Unsupported environments or clipboard failures produce an explanatory message.

verified means an explicitly authorized live native tool-call/tool-result E2E was recorded; no providers are checked in as verified by default.

MCP

MCP is built in. Put a version 1 mcp.json beside your model config.json, or in .laohuang/mcp.json under the directory where you start the CLI:

{
  "version": 1,
  "servers": {
    "local": {
      "transport": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/mcp-server.js"]
    }
  }
}

Transports are stdio, Streamable http, and legacy sse. Use /mcp status, /mcp reload, /mcp reconnect <server>, /mcp login <server>, and /mcp logout <server>. Remote OAuth uses auth: { "type": "oauth" }. Environment references use { "env": "VARIABLE_NAME" }; no shell expansion occurs. Project entries replace same-name global entries entirely. Invalid files are skipped on startup; an invalid reload keeps the previous configuration.

Connections start in the background; the first model request waits for initial discovery. Below 20 available tools, definitions are provided in full. At 20 or more, tool_search automatically searches and loads MCP tools; read, write, edit and bash remain directly available. Search is local BM25 with query and limit (default 8, range 1–20). Loaded definitions persist in retained session history; after compaction removes them, search loads them again. The model can start OAuth through a synthetic authenticate tool; the terminal shows a URL immediately and waits up to 15 minutes for the user to finish authorization. Tool calls default to sequential execution. Recoverable network errors allow at most three application-level attempts; lost responses can cause duplicate side effects. Cancellation, timeouts and business errors are not retried. Large outputs and media are retained under mcp-artifacts beside the global configuration; returned paths let you retrieve them. Media files are not native image/audio model input. MCP processes run with your user permissions and are not sandboxed.

See the repository README for development and architecture details: https://github.com/hxr223/laoHuangCode#readme