Skip to main content
The AgentMark agent skill teaches AI coding agents how to use AgentMark: CLI commands, .prompt.mdx authoring, datasets, experiments, evals, and git-based deploys. Once installed, a compatible agent auto-discovers it at session start and applies it whenever AgentMark context appears: .prompt.mdx files, agentmark.json, @agentmark-ai/* imports, or an explicit mention of AgentMark.
This page covers the agent skill (workflows and conventions your agent follows). It complements the docs MCP, which lets your editor query AgentMark documentation, and the gateway MCP, which lets it drive the full AgentMark API: traces, datasets, scores, experiments, apps, and deployments.

​
Installed automatically with create agentmark

If you bootstrap a project with the scaffolder, you already have the skill, with no extra steps: 
npm create agentmark@latest
The scaffolder installs the skill to ./.agents/skills/agentmark/ and symlinks it into each tool’s native location (Claude Code, Codex, Cursor, GitHub Copilot, and other agentskills.io-compatible tools).

​
Adding the skill to an existing project

Pick whichever fits your tooling:
Works across Claude Code, Codex, Cursor, GitHub Copilot, Amp, and other agentskills.io-compatible tools:
npx skills add agentmark-ai/skills
Files land in ./.agents/skills/agentmark/ with symlinks into per-tool paths created automatically.
After installing, start a new agent session in your project. The skill is discovered automatically; there is nothing to enable.

​
What the skill contains

skills/agentmark/
├── SKILL.md                     # entry point, read by the agent on auto-invoke
├── workflows/                   # one file per common task
│   ├── building-datasets.md
│   ├── creating-prompts.md
│   ├── deploying.md
│   ├── headless-with-mcp.md
│   ├── observability.md
│   ├── running-experiments.md
│   ├── setup-and-integration.md
│   └── using-evals.md
└── reference/                   # auto-generated from source
    ├── cli-commands.md
    └── frontmatter-schema.md

​
How it stays current

The skill is a thin pointer that delegates to live sources rather than a static snapshot. It teaches agents to:
  1. Run npx @agentmark-ai/cli <cmd> --help for the canonical CLI surface
  2. Fetch docs.agentmark.co/llms.txt for documentation navigation
  3. Fetch specific pages from these docs for current content
The bundled reference/ files are regenerated from source on every release, so a stale install degrades gracefully: the agent is always instructed to prefer the live CLI and docs over the bundled copies.

​
Related documentation

Quickstart

Bootstrap a project with npm create agentmark, skill included

AI editor integration

Wire the docs MCP so your editor can query AgentMark documentation

Gateway MCP

Drive the full AgentMark API from your editor: traces, datasets, scores, deployments

CLI commands

The command surface the skill drives
Next: Client setup shows what the skill scaffolds when you ask it to set up AgentMark (the client, the dev entry, and the handler).

​
Have questions?

Reach out any time: