dt-ai-toolkit 0.3.3

Data-team toolkit for building, interacting with, and improving data projects with coding agents

dt-ai-toolkit

A CLI that gets a whole data team working with coding agents the same way.

If your team uses Claude Code for data work, every teammate ends up solving the same problems by hand: how to lay out a project so an agent can work in it, which skills to give the agent, how to run repeatable agent workflows, and how to turn analysis into a polished PDF. dt-ai-toolkit packages those answers into one installable tool — run dat setup in a fresh directory and a teammate gets the same agent-ready project, curated skills, and report pipeline as everyone else.

What it does

• Scaffolds agent-ready data projects — standard folders (data/, reports/, sandbox/), a managed .gitignore, and tool installs, customizable per team via a TOML profile.
• Keeps reusable AGENTS.md templates — house rules stored once in the toolkit database with {placeholder} gaps, rendered into a project as AGENTS.md plus a generated CLAUDE.md that imports it, so there is one source of truth instead of two copies that drift (dat agent-markdown).
• Installs curated Agent Skills — a bundled set focused on data honesty (profiling datasets before trusting them, critiquing charts and talking points), plus install from any local folder or git repo, with provenance tracking.
• Runs agent workflows — repeatable flows via the Claude Agent SDK, defined as plain markdown files with YAML frontmatter. Bundled flows included; write your own in minutes.
• Generates real reports — scaffolds a React TSX → HTML → PDF pipeline so agent output can ship as a formatted document, not a wall of markdown.
• Evaluates skills & agents in a gym — run curated tasks against a skill or agent in an isolated sandbox, score the transcript with an LLM judge, or A/B two targets head-to-head (dat gym run / dat gym ab / dat gym manage). Makes real API calls.
• Chat-edits everything with the agent of your choice — every record in the workshop can be built or edited in an interactive coding-agent chat (multi-line block paste, split-screen live preview, /copy to clipboard), backed by Claude (pick model + effort per session, up to fable) or any CLI coding agent you register (dat providers).
• Puts it all in one terminal app — dat workshop is a single Textual UI over every record, with ten workspaces on a grouped rail: docs, then setup / agent-markdown / skills / agents / tools / providers / harnesses, then the gym and benchmarks. alt+1…alt+9 and alt+0 jump, ctrl+k finds any record by name, and ? shows every key the current screen answers to. Every dat <group> manage command is a deep link into it.
• Keeps everyone unblocked — dat doctor checks the whole environment (uv, git, claude CLI, auth, bun, playwright) and dat docs is a built-in interactive docs browser.

Agents, skills, and setup profiles live in a local SQLite database, so teams can inspect, edit, back up, and share their configuration (dat db info, dat agent backup / restore, and the same for skills and profiles).

Install

# no install needed — run it straight from PyPI
uvx dt-ai-toolkit --help

# or install it once and get the short `dat` command
uv tool install dt-ai-toolkit

# prefer not to install? alias `dat` to uvx in your shell config instead
uvx dt-ai-toolkit alias

Requires Python 3.11+ and uv. Every dat ... command below also works as uvx dt-ai-toolkit ....

Quickstart

dat db setup       # one-time: create the toolkit database, seed bundled content

mkdir my-project && cd my-project
dat setup          # folders + gitignore + installs Claude Code & omnigent
dat doctor         # verify the environment
dat workshop       # the interactive app over everything (opens on the docs)

Commands

command	what it does
workshop	The interactive app over every record: ten workspaces on a grouped rail (docs · setup/agent-markdown/skills/agents/tools/providers/harnesses · gym/benchmarks). dat workshop [workspace] [--select NAME]; every manage command deep-links into it.
setup	Scaffold a data project (data/, reports/, sandbox/, managed .gitignore block) and install tooling. Customizable via a dt-setup.toml profile (setup config init).
agent-markdown	Reusable AGENTS.md templates as database records (frontmatter markdown with {placeholder} defaults); renders AGENTS.md plus a generated CLAUDE.md that imports it. list / restore / backup / manage (alias: agentmd).
skills	Install Agent Skills into .claude/skills/ from the curated set, a local folder, or a git repo. list / install / update with provenance tracking.
agent	Run agent flows via the Claude Agent SDK: bundled sourcing-report and project-review, or your own frontmatter-markdown agent files. agent help <name> shows an agent's docs; run --directory picks the working directory.
report	Scaffold the React TSX → HTML/PDF report pipeline (report new <name>).
tools	Register plain-Python tools as harness-specific wrappers — an MCP server for Claude, a CLI + skill for pi (tools register <name> --for claude|pi).
gym	Evaluate, run, and A/B test skills & agents against curated tasks with an LLM judge (gym run / gym ab / gym manage).
benchmark	Run a saved gauntlet across harness/model combos and compare them in a task × combo matrix (benchmark run / results, benchmark harness add).
db	Manage the toolkit database — the source of truth for agents, skills, and setup profiles. setup / info / path / upgrade-backup.
providers	Configure which agents back the interactive chats: claude built in, plus any coding-agent CLI (opencode, pi, ...) via providers.toml — list / init / manage.
doctor	Check the environment: uv, git, claude CLI, omnigent, auth, bun, playwright.
alias	Write a dat alias into your shell config (bash, zsh, fish, PowerShell) so dat works without a tool install.
docs	Browse documentation — an interactive TUI, or --plain / show <topic> / export for non-interactive use.

Singular and plural are interchangeable: setup/setups, skill/skills, agent/agents, and gym/gyms all resolve to the same command.

Bundled skills

The curated skills lean toward data honesty — making sure what an agent (or you) produces can survive scrutiny:

• interrogate-data — structured protocol for profiling and stress-testing a dataset before trusting it
• critique-talking-points — adversarial claim-by-claim review of narratives before they ship
• critique-visualization — chart-honesty review, including re-deriving plotted values from source data
• tsx-report-generator — walks Claude through the full TSX → HTML/PDF report pipeline
• dt-setup-config — helps Claude build a custom dt-setup.toml for your team
• dt-ai-toolkit — meta skill teaching Claude how to call this CLI itself (uvx dt-ai-toolkit ...)
• grill-me — Socratic quiz rounds that test your understanding of a dataset, pipeline, or report before someone else does

dat skills list                        # the toolkit-db catalog (what you install from)
dat skills install                     # install every db skill (idempotent sync)
dat skills install interrogate-data    # a single db skill by name
dat skills installed                   # what's installed in this project
dat skills install https://github.com/your-org/team-skills.git --all

Custom agent flows

Agents are markdown files with YAML frontmatter — config on top, prompt below, {placeholders} filled from the command line. A one-line description and a longer help doc (what it does and when to use it, shown by dat agent help) are both required:

---
name: data-audit
description: Audit a dataset for quality issues
help: |
  Profiles a dataset and reports quality issues — nulls, outliers, type
  mismatches — with the query behind every finding. Point --arg data_path at
  the file to audit.
allowed_tools: [Read, Glob, Grep, Bash]
args:
  data_path: {required: true}
---
Audit the dataset at {data_path}. Cite the query behind every finding.

dat agent help data-audit.md                                   # read its docs first
dat agent run data-audit.md --arg data_path=data/sales.csv
dat agent run sourcing-report --arg topic="Q3 readmissions" --dry-run
dat agent run data-audit.md --arg data_path=data/sales.csv --directory ../warehouse

Development

uv sync
uv run pytest
uv run dat --help

See CLAUDE.md for architecture conventions.