Model Locally with the CLI

cred init sets up a directory for local Malloy modeling. It logs you in, connects the directory to a Credible environment, installs the modeling skills, and starts a local Malloy Publisher serving your model. Your local Publisher sends its queries through Credible, so it reads the same warehouse your published models do. No warehouse credentials are stored on your machine.

It works with any coding agent that reads a project directory: Claude Code, Codex, Cursor, Gemini CLI and others.

Prerequisites

• Node.js 20+ and the CLI (npm install -g @credibledata/cred-cli)
• A Credible organization, and an environment whose connections your admin has set up. See Connect a Database

Get Started

Run it in an empty directory, or in an existing Malloy package:

mkdir my-model && cd my-model
cred init

cred init asks for anything it doesn't know yet:

1. Log in, if you aren't already.
2. Pick an organization and environment. The choice is saved in the directory, so later commands there use the same pair.
3. Write the workspace files (listed below), including a local proxy for each of the environment's connections.
4. Install the modeling skills.
5. Start the local Publisher at http://localhost:4000 and open it in your browser. Edits to .malloy files reload as you save.

The first run downloads the Publisher, which takes about 30 seconds. If the environment has no connections yet, the Publisher still starts, but your model has nothing to query. Ask an admin to add a connection in Credible, then run cred dev again.

Press Ctrl-C to stop the Publisher. To start it again later, run this in the same directory:

cred dev

cred dev reuses the saved organization and environment, refreshes the skills, and starts the Publisher in the foreground.

What It Writes

In a Git repository, cred init adds .credible/ and the skill folders to .gitignore, because they belong to you rather than to the repository. cred publish leaves every file in this table except publisher.json out of the package.

Skills

cred init installs Credible's open-source skills. They go in two places:

• .claude/skills/, where Claude Code finds them
• .agents/skills/, for Codex, Gemini CLI and other agents

It also writes a short rule file, .claude/rules/credible-modeling.md, that tells the agent to read the skills before it starts modeling. If the directory has a .cursor/ folder, it writes a Cursor rule as well.

Which skills: the modeling-ide set, the same set the VS Code extension installs. The Credible deployment you're signed in to decides which version you get, so everyone on that deployment gets the same skills. cred status --json shows the versions installed in a directory.

When skills update: every cred init and cred dev checks for a newer version of the skills. When there is one, cred replaces the files it installed last time. It never touches skills it didn't install, except as described under Keep below. When the skills are already current, nothing is written. If the skills can't be fetched, the Publisher starts anyway and the next cred dev tries again. Fully offline, cred init and cred dev stop earlier, with exit code 1, because they read the environment's connections from Credible before starting the Publisher.

Skills that were already there: a directory made with npm create @malloy-publisher/malloy-package comes with its own skills. The first time, cred init asks what to do with them:

• Replace (recommended): moves them to .credible/replaced-skills/ and installs Credible's set. Credible's set includes the same Malloy skills. Nothing is deleted.
• Keep: installs Credible's set alongside them. A skill with the same name as one of Credible's is overwritten, and a later update removes it. Skills from npm create @malloy-publisher/malloy-package share names with Credible's, so choose Replace for those.
• None: installs no skills, and later runs remember that. Run cred dev --skills replace or cred dev --skills keep to start installing them.

Bring Your Own Connections

If the directory already has a publisher.config.json from a local Publisher setup, cred init offers to copy its warehouse connections into your Credible environment. Credible tests each connection before adding it, and a connection that fails its test is not added. If you named connections with --connections and any of them fails its test, none are added, so you can fix the list and run it again. A connection whose name the environment already has is not replaced: change it with cred update connection instead. DuckDB connections aren't copied: publish their files with the package instead.

Use It from an Agent or Script

Every question cred init asks has a flag. Without a terminal, cred init asks nothing. If it's missing an answer, it exits with code 2 and names the flag to pass.

# A new environment, no connections to copy, Publisher not started yet
cred init -o acme -e my-env --create-environment --connections none --no-start

# Start the Publisher in the background, then wait until it serves the package.
# -y skips the "Use organization...?" question in an interactive shell.
# The first run downloads the Publisher, so this can take about 30 seconds.
cred dev --no-open -y &
until cred status --json | jq -e '.publisher.ready' > /dev/null; do sleep 2; done

Publish

When the model is ready, publish it from the same directory:

cred publish --dry-run   # list what would be uploaded, and upload nothing
cred publish

Besides the workspace files above, the upload leaves out credential files and anything listed in a .credignore file. See the CLI reference for the full list, and Publishing for versions and setting the latest version.