Installation

The archmax harness requires Node.js 22 or later and ships as an ESM-only npm package.

Install the package

npm install @archmax-ai/harness

This installs the runtime library and the archmax CLI.

For hacking on the SDK itself, see development setup.

The authoring skill

The authoring skill is the Agent Skill that coding agents (Claude Code, Cursor, Codex, …) use to author workflow.yaml workspaces and wire the runtime. Install it with the skills CLI:

npx skills add archmax-ai/harness

That installs the latest version from the GitHub repository. Add -g to install it for your user rather than the current project.

The package also ships the skill inside the install, so it always matches the SDK version you depend on. After npm install, install the packaged copy:

npx skills add ./node_modules/@archmax-ai/harness

npx skills experimental_sync does the same for every skill-bearing package in node_modules. The package ships the skill at skills/archmax-harness/ for these installers.

To load it in your own agent, point skills middleware at the exported constant:

import { BUNDLED_AUTHORING_SKILL_DIR } from "@archmax-ai/harness";
// <BUNDLED_AUTHORING_SKILL_DIR>/archmax-harness/SKILL.md (+ references/)

Configure the model

The model is configured from environment variables, and it targets any OpenAI-compatible endpoint (OpenRouter, vLLM, LiteLLM, Together, …). These five are what a first run needs:

Variable	Description	Example
ARCHMAX_API_BASE_URL	Base URL of the OpenAI-compatible API	https://openrouter.ai/api/v1
ARCHMAX_API_KEY	API key for the endpoint	sk-or-v1-...
ARCHMAX_MODEL	Model identifier	anthropic/claude-sonnet-4.6
ARCHMAX_TEMPERATURE	(optional) sampling temperature	0
ARCHMAX_MAX_TOKENS	(optional) max output tokens	4096

Under the hood this builds a ChatOpenAI instance pointed at ARCHMAX_API_BASE_URL, and Deep Agents uses it as its model.

Prompt caching, streaming and token pricing have their own variables. The configuration reference lists the full set.

.env is loaded from the workspace root. That is the directory passed via --root (or rootDir), or the current working directory. Variables already present in the real process environment always win.

cp .env.example .env   # then fill in your credentials

To supply a model programmatically instead, see the public API reference. Any LangChain BaseChatModel works, including per-role factories.