Installation

You need Python 3.10 or later, uv on your PATH, and one of the supported coding agents (Claude Code, Cursor, or Codex) already installed.

Add the AI Harness to a project

Run the following command in the directory where you want the workspace:

uvx dlthub-init@latest

This scaffolds a dltHub workspace, then wires the base init toolkit for Claude Code (default):

your-project/
├── .dlt/
│   ├── .workspace              # marks this directory as a dltHub workspace
│   ├── config.toml             # workspace-wide config
│   └── secrets.toml            # workspace-wide secrets (gitignored)
├── .agents/
│   └── skills/                 # vendored source of the `init` toolkit
│       ├── dlthub-router/
│       ├── setup-secrets/
│       └── improve-skills/
├── .claude/
│   └── skills/                 # symlinks to .agents/skills/*
├── .mcp.json                   # registers `dlt-workspace-mcp`
├── __deployment__.py           # empty; jobs get declared here
└── pyproject.toml

To wire up Cursor or Codex instead, run:

uv run dlthub ai init --agent cursor

Swap cursor for codex as needed.

What you get

Your workspace now contains the base init toolkit, which ships an MCP server (dlt-workspace-mcp), a workspace-setup rule, and three skills:

• dlthub-router: routes user intent to the right toolkit and installs it if missing.
• setup-secrets: safely manages .dlt/secrets.toml without exposing values to the agent.
• improve-skills: captures new patterns learned in a session so skills stay lean.

Feature toolkits (REST API pipelines, SQL database, transformations, deployment, and so on) are not installed by default. dlthub-router installs them automatically when it matches your intent to one, or you can install them manually from the CLI.

Adding feature toolkits

Most of the time you don't install feature toolkits manually. You talk to your agent about what you want to build, and dlthub-router picks the right toolkit and installs it for you.

For example, if you tell the agent:

I want to ingest pull requests and issues from the GitHub REST API.

dlthub-router matches that intent to the rest-api-pipeline toolkit, runs dlthub ai toolkit install rest-api-pipeline under the hood, and hands off to that toolkit's entry skill (find-source).

If you prefer explicit installs, the same three CLI commands are always available. List everything:

uv run dlthub ai toolkit list

Inspect a specific toolkit before installing:

uv run dlthub ai toolkit info rest-api-pipeline

Install it:

uv run dlthub ai toolkit install rest-api-pipeline

By default dlthub ai toolkit install installs for the agent already wired up in the workspace. Pass --agent claude|cursor|codex to install for a different one, or --overwrite to replace files the agent already has.

Verify

Check that the workspace is fully wired:

uv run dlthub ai status

Output includes the dlt version, the configured agent, the installed toolkits, and warnings when the MCP server or any dependency is missing. Fix the warnings before you start working with your agent.

What's next

• Toolkits walks through the full catalog and where each fits in the development cycle.
• Deploy with AI Harness shows how to use the dlthub-platform toolkit end-to-end.