dbt Agent Skills - Setup Guide¶

dbt Labs publishes its own Agent Skills collection for working inside dbt projects: build and modify models, write tests, create MetricFlow semantic models, govern a Mesh deployment, troubleshoot platform job failures, and migrate a project from dbt Core to dbt v2. The repo is the publisher's own, not a community mirror, and it is versioned with a changie changelog and GitHub Actions CI.

The design choice that matters is that the skills are not slash commands. Fourteen of the 17 SKILL.md files set user-invocable: false, so the agent loads the relevant skill when a prompt matches its description and the operator rarely has to remember a command name. The repo also ships its own evaluation harness in evals/, an A/B tool that compares skill variations against recorded scenarios, and a changie changelog with CHANGELOG.md.

Source: dbt-labs/dbt-agent-skills via skills.sh - 12.4K combined installs across 19 indexed listings; first seen Jan 29, 2026 GitHub: dbt-labs/dbt-agent-skills (746 stars; Apache-2.0; branch main, pushed 2026-10-22; 17 SKILL.md directories: 12 under skills/dbt/skills/, 3 under skills/dbt-migration/skills/, 1 under skills/dbt-extras/skills/, plus an unindexed maintainer skill at .claude/skills/auditing-skills/SKILL.md) Category: Data / Analytics Engineering Quality Tier: 🟢 Production - official dbt Labs repo under Apache-2.0, 16 of the 19 indexed listings verified against the tree, a pytest suite plus CI, and a real A/B skill-evaluation harness in evals/; the flagship skill carries a Snyk warn, noted under Security

Installation¶

# Preview, then install every skill in the repo
npx skills add dbt-labs/dbt-agent-skills --list
npx skills add dbt-labs/dbt-agent-skills

# Install only the everyday analytics engineering skills
npx skills add dbt-labs/dbt-agent-skills/skills/dbt

# Install only the one-off migration skills
npx skills add dbt-labs/dbt-agent-skills/skills/dbt-migration

# One skill, or a global install into ~/.<agent>/skills/
npx skills add dbt-labs/dbt-agent-skills --skill using-dbt-for-analytics-engineering
npx skills add dbt-labs/dbt-agent-skills --global

# Claude Code plugin path, from the README
/plugin marketplace add dbt-labs/dbt-agent-skills
/plugin install dbt@dbt-agent-marketplace
/plugin install dbt-migration@dbt-agent-marketplace

# Or install as dbt packages, which also works for dbt Cloud and the dbt Wizard CLI
packages:
  - git: "https://github.com/dbt-labs/dbt-agent-skills.git"
    subdirectory: "skills/dbt"
    revision: main

# Opt in via dbt_project.yml flags (ai_provider: claude writes to .claude/skills)
dbt deps

Prerequisites from the README: dbt installed and configured, a project with dbt_project.yml, and basic familiarity with models, tests and sources. fetching-dbt-docs and configuring-dbt-mcp-server work without an existing project. The platform skills read a DBT_TOKEN from the environment rather than a literal token in config.

What It Provides¶

Skill	Installs	Use For
using-dbt-for-analytics-engineering	1,169	The default dbt skill: build and modify models, write SQL with ref() and source(), add tests, debug errors, explore unfamiliar sources, validate with dbt show
fetching-dbt-docs	1,105	Retrieve and search dbt documentation in LLM-friendly markdown, including an llms.txt page index and a script that caches the full docs locally
adding-dbt-unit-test	1,103	Write unit test YAML that mocks upstream inputs and asserts outputs, with per-warehouse type and caveat references for BigQuery, Snowflake, Redshift, Postgres and Spark
running-dbt-commands	1,059	Get the right dbt executable, flags, selectors and parameter formats for run, build, test, compile and show
answering-natural-language-questions-with-dbt	1,035	Answer business questions (revenue, KPIs, top customers) by querying the semantic layer or warehouse SQL, not by editing models
configuring-dbt-mcp-server	1,032	Generate MCP server config, resolve authentication, and validate connectivity for Claude, Cursor or VS Code
building-dbt-semantic-layer	1,023	Create semantic models, metrics, dimensions, entities and time spines with MetricFlow, covering latest and legacy YAML specs
troubleshooting-dbt-job-errors	1,014	Diagnose dbt platform job failures from run logs, the Admin API and git history, with an explicit iron rule against editing a test to make it pass
working-with-dbt-mesh	869	Judge whether a column rename, removal or retype is breaking, then handle model versions, contracts, access, groups and cross-project refs
using-dbt-state	554	Configure and debug dbt State (server-backed node reuse) and separate it from the state:modified selector and --state deferral
creating-mermaid-dbt-dag	527	Generate a Mermaid flowchart of model lineage from the dbt MCP lineage tools, manifest.json, or direct code parsing as a fallback
migrating-dbt-project-across-platforms	526	Move a project between warehouses (Snowflake to Databricks and back) using Fusion's real-time compilation to find SQL dialect differences

The remaining 7 indexed listings range from 56 to 385 installs. Four of them are documented in the repo: maintaining-dbt-documentation (329), migrating-dbt-core-to-v2 (182), upgrading-dbt (129) and querying-the-dbt-information-schema (128). Three carry install counts without a matching directory in the current tree: migrating-dbt-core-to-fusion (385), upgrading-dbt-core (132) and using-dbt-index (56).

Why This Matters for Hermes Agents¶

dbt work is command-line work with exact syntax: which executable, which selector, which flag order, whether a variable is a string or a dict. running-dbt-commands encodes those formats so an agent stops inventing flags, and fetching-dbt-docs gives it a way to check a reference page instead of answering from memory, including a local 24-hour cache of docs.getdbt.com/llms-full.txt.

The second gain is judgement, not syntax. working-with-dbt-mesh is built around one question an agent usually gets wrong: is this change breaking, and who consumes the model. troubleshooting-dbt-job-errors carries an explicit iron rule (never modify a test to make it pass without understanding the failure) and a table of rationalizations that mean stop. Those rules are exactly the ones a helpful-sounding agent skips under time pressure.

The semantic layer skills close the gap between a business question and a metric definition. building-dbt-semantic-layer covers metric types and both the latest and legacy YAML specs, so an agent can tell which spec a project is on before editing, and answering-natural-language-questions-with-dbt is scoped to querying rather than to changing models, which keeps a read-only question from turning into a schema edit.

Finally, the repo is testable like software. Everything follows the Agent Skills specification, CI validates frontmatter and links, and evals/ A/B tests skill variations against recorded scenarios. For an operator maintaining a fleet of agents, that means a skill change can be measured rather than argued about.

Usage¶

You say	What happens
"Add a new incremental model for subscription revenue"	Loads using-dbt-for-analytics-engineering, plans the model against real sources, writes the SQL with ref(), then runs dbt show to validate
"Add unit tests for the orders model"	Loads adding-dbt-unit-test and writes YAML that mocks upstream rows and asserts outputs, with the warehouse-specific caveats for the project's adapter
"Build a revenue metric in the semantic layer"	Loads building-dbt-semantic-layer, checks whether the project is on the latest or legacy spec, then writes the semantic model, measure and metric
"What were sales last quarter?"	Loads answering-natural-language-questions-with-dbt and queries the semantic layer or warehouse instead of editing any model
"This nightly job failed and the error is vague"	Loads troubleshooting-dbt-job-errors, pulls run logs and Admin API data, checks recent git history, and reports the root cause before any fix
"We are renaming a column used by three dashboards"	Loads working-with-dbt-mesh and returns a breaking-change assessment plus a versioning or migration path
"Set up the dbt MCP server for Cursor"	Loads configuring-dbt-mcp-server, generates the config with `${DBT_TOKEN}` references and validates connectivity
"Draw the lineage for the finance models"	Loads creating-mermaid-dbt-dag and emits a Mermaid flowchart from the lineage tools or manifest.json

Verification¶

Confirm what the skills CLI installed:

npx skills list | grep -ci dbt

Read the skills before installing. The seven URLs below returned 200 with curl -s -o /dev/null -w "%{http_code}" when checked on Oct 26, 2026:

curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/README.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt/skills/using-dbt-for-analytics-engineering/SKILL.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt/skills/running-dbt-commands/SKILL.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt/skills/fetching-dbt-docs/SKILL.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt/skills/troubleshooting-dbt-job-errors/SKILL.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt-migration/skills/upgrading-dbt/SKILL.md
curl -sL https://raw.githubusercontent.com/dbt-labs/dbt-agent-skills/main/skills/dbt-extras/skills/creating-mermaid-dbt-dag/SKILL.md

All three skills directories nest one level deeper, at skills/<area>/skills/<name>/SKILL.md, so a plain skills/<name>/SKILL.md fetch will 404 on this repo.

Security¶

skills.sh verdicts for sampled skills (verified Oct 26, 2026) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
using-dbt-for-analytics-engineering	Pass	Pass	Warn
running-dbt-commands	Pass	Pass	Pass
fetching-dbt-docs	Pass	Pass	Pass

Permission grants are declared per skill through allowed-tools, and only five SKILL.md files declare one. using-dbt-for-analytics-engineering grants Bash(dbt *), Bash(jq *), Read, Write, Edit, Glob and Grep, which is real write access to the project: the agent can edit models, tests and YAML and can run dbt run or dbt build against the configured warehouse, spending compute and changing tables. maintaining-dbt-documentation adds Bash(python3 *) and Bash(git *). The two migration skills that declare grants allow Bash(dbt:*), Bash(git:*) and Bash(uvx:*), and migrating-dbt-core-to-v2 additionally grants WebFetch(domain:api.github.com) to search dbt-fusion issues. creating-mermaid-dbt-dag grants only the two dbt MCP lineage tools, Read, Glob, Grep and Bash(jq *).

The other twelve skills declare no allowed-tools at all, so whatever your host agent already permits is the ceiling. Network behaviour: fetching-dbt-docs downloads https://docs.getdbt.com/llms-full.txt into ~/.cache/dbt-docs with a 24-hour cache, and configuring-dbt-mcp-server writes MCP config files that reference ${DBT_TOKEN} and explicitly warns against pasting literal tokens into committed config. troubleshooting-dbt-job-errors needs a dbt platform service token to reach the Admin API. The single Snyk warn sits on the flagship skill, which is the one with both write access and dbt * shell scope, so review it against your own blast radius before enabling it on a production project.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Most skills assume a working dbt installation and an existing project. Only fetching-dbt-docs and configuring-dbt-mcp-server stand alone.
Three indexed listings have no directory in the current tree: migrating-dbt-core-to-fusion (385), upgrading-dbt-core (132) and using-dbt-index (56). They are install counts without a fetchable SKILL.md at this snapshot.
using-dbt-state (554) is in the tree but absent from the README's Available Skills table, which documents 14 of the 17 skills.
The migration skills are one-off by the README's own description. Installing the whole repo for every session loads context that only pays off during an upgrade.
14 of the 17 skills set user-invocable: false, so routing is by prompt match rather than a command name. The exceptions are upgrading-dbt, migrating-dbt-core-to-v2 and the internal auditing skill.
The docs skills fetch live documentation, so answers track docs.getdbt.com rather than the dbt version pinned in your project.
Snapshot data, verified Oct 26, 2026: 12,357 combined installs across 19 indexed listings; 746 GitHub stars; Apache-2.0; last pushed 2026-10-22. Counts drift over time.

Databricks Agent Skills - Setup Guide - watch the warehouse side of a dbt project that runs on Databricks.
Amplitude Agent Skills - Product Analytics Setup - the downstream analytics layer that consumes the models these skills build.
PlanetScale Database Skills - Setup Guide - schema and query work at the database layer beneath a dbt source.
Redis Agent Skills - Official Data Modeling & Caching Setup - data modeling conventions for a second data store alongside the warehouse.
Skills Catalog | Skills Marketplace

Pro Tips¶

Install skills/dbt globally and keep skills/dbt-migration scoped to the migration project. The README says the migration set is one-off, and it is 3 of the 16 available skills.
Read allowed-tools in the SKILL.md before enabling the analytics engineering skill on a production project. It can edit models and run builds, which is a different risk class from a read-only docs skill.
Point the agent at using-dbt-state when someone reports models rebuilding that should have been reused. The skill exists specifically because dbt State gets confused with state:modified.
Use working-with-dbt-mesh before a column rename rather than during the incident. It is designed to answer who consumes the model while the change is still optional.
Keep the docs cache warm. fetching-dbt-docs reuses ~/.cache/dbt-docs for 24 hours, so repeated lookups in a long session do not refetch llms-full.txt.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ