Install · 30 seconds · free · expert skills

Add WinDAGs to the tool
you already use.

Your agent gets a library of expert skills and a way to plan work before it runs. Start simple; the technical install paths are below.

brew install →No Homebrew? →What does this install? →
✓ Claude Code✓ Claude Desktop✓ Cursor✓ Codex CLI✓ Gemini CLI✓ Any MCP client
▶ Live · this is the actual install30 sec
Terminal recording of installing windags-skills via the Claude plugin marketplace

What you'll have when you're done

  • Your AI agent can search and use specialist skills when the work needs real domain knowledge.
  • In Claude Code: a /next-move slash command that predicts your next 3-wave action plan from your git state and recent files.
  • No API keys for the read-only install path. Search and skill lookup run locally; execution uses whatever provider you configure in your client.

Prerequisites

  • One of: Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI, or any MCP-compatible client.
  • Node.js 18 or later (only for the MCP path; Claude Code path doesn't need it).
  • About 30 seconds.

One command. Every tool you have.

windags init auto-detects Claude Code, Claude Desktop, Cursor, Codex, and Gemini CLI on your machine and wires them all in one shot — skills, agents, slash commands, and the MCP server.

shell
$ brew install curiositech/windags/windags && windags init

macOS / Linux (with Homebrew). On Windows or WSL without Homebrew, scroll to Path 1 for the Claude Code plugin marketplace install.

Path 1

Plugin marketplace (Claude Code)

The fastest install. Adds the marketplace, installs the plugin, and you're done. Slash commands and skills appear automatically in Claude Code.

shell
$ claude plugin marketplace add curiositech/windags-skills && claude plugin install windags-skills
What lands
551 skills + 5 meta-DAG subagents in ~/.claude/skills/ and ~/.claude/agents/.
What it costs
$0 to install. Skills run on whatever model your client already uses.
Verify
In any Claude Code session, type /next-move. If it autocompletes, you're set.
Path 2

MCP server (any client)

The MCP server exposes 9 tools over stdio: the 4 retrieval primitives (windags_skill_search, windags_skill_graft, windags_skill_reference, windags_history) plus 5 planner-grade tools added in v2.8.0: batch search and graft, windags_node_requirements (provider-native model IDs, no more bare "haiku"/"sonnet" leaking to non-Anthropic providers), windags_validate_dag for schema-checking, and windags_estimate_cost for planning-time cost surfacing. Full reference at /tools. Zero API keys for the read-only tools — the cascade runs locally.

Claude Code

shell
$ claude mcp add windags -- windags-mcp

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

json
{
  "mcpServers": {
    "windags": {
      "command": "windags-mcp"
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json:

json
{
  "mcpServers": {
    "windags": {
      "command": "windags-mcp"
    }
  }
}

Codex / Gemini CLI / generic stdio MCP

Any client that supports stdio MCP works. After brew install curiositech/windags/windags, point your client's MCP config at windags-mcp over stdio.

Verify

  1. Open a session in your client and ask: "search WinDAGs for skills about Postgres connection pooling". The agent should call windags_skill_search and return ranked matches.
  2. Then: "graft the top match into context". The full skill body should appear and the agent's next answer should be visibly more specific.
  3. For Claude Code only — type /next-move to invoke the prediction pipeline.

Privacy

By default the MCP server fires one fire-and-forget POST to api.windags.ai/v1/events per machine per 24 hours, carrying a hashed machine ID (sha256(hostname+username), truncated to 16 hex chars), the tool name, plugin version, Node major, and OS platform. No raw task text. A small ~/.windags/telemetry-state.json file tracks the last-sent timestamp for sampling. Set WINDAGS_TELEMETRY=off for zero network calls; WINDAGS_TELEMETRY=anonymous (the default) sends the hashed signal above; WINDAGS_TELEMETRY=full opts in to sending the raw task text so we can see where the cascade misses.

Troubleshooting

Slash commands don't appear after plugin install
Restart Claude Code. The plugin loader scans on startup.
MCP tools don't appear in the picker
Confirm the JSON config validates. Restart the client. In Claude Code, run claude mcp list to confirm the server is registered.
windags-mcp: command not found
The brew formula puts windags-mcp in /opt/homebrew/bin (Apple silicon) or /usr/local/bin (Intel). If your shell PATH doesn't include those, run `brew --prefix windags` to find the install location and use the absolute path to bin/windags-mcp.
No Homebrew (Linux / Windows / WSL)
Fall back to Path 1 (plugin marketplace) — it includes /next-move and the full catalog. For MCP without brew, clone curiositech/windags-skills, run `npm install` in mcp-server/, and point your client config at the absolute path to mcp-server/index.js.
Tool2Vec embeddings missing / cache cold
The first graft generates Tool2Vec data on demand and caches it under ~/.windags/. Subsequent grafts are fast.
Ready when you are

Run one command. Keep what helps. Uninstall in one second if it doesn't.

One brew command, every AI tool you have wired in. Uninstall is brew uninstall windags. Hashed telemetry is on by default — disable any time with WINDAGS_TELEMETRY=off.

shell
$ brew install curiositech/windags/windags && windags init
← Read the blogBrowse 551 skills →GitHub: windags-skills →