Author with Claude Code

A coding agent can write your semantic model YAML. hashictl gives it a safe way to do that. The agent never writes the Draft. It opens a Proposal, and an editor approves it in Studio.

login -> model pull -> (agent edits) -> validate -> test -> diff -> propose
                                                                     |
                                                 editor approves in Studio

Install the skills and the CLI

In Claude Code, add the Hashi plugin:

/plugin marketplace add datahashi/agent-skills
/plugin install hashi@datahashi

For Cursor or another agent, copy the folders under skills/ from datahashi/agent-skills to where your agent reads skills.

The skills call hashictl. Download it from the Releases of the same repository: hashictl_<tag>_<os>_<arch>.tar.gz (.zip on Windows), with a checksums file. Put the binary on your PATH.

Sign in

hashictl login --url http://localhost:8080 --email you@example.com

Hashi sends a login code to your email. In a local community install the code is in the server log. The CLI stores the session in ~/.config/hashictl/config.json, readable only by you. For an agent running in CI, set HASHI_URL, HASHI_TOKEN and HASHI_TENANT instead.

Change the model

hashictl model pull --dir model     # copy the Draft into ./model
# ask your agent to edit model/entities/order.yaml
hashictl validate --dir model       # compile on the server; writes nothing
hashictl test --dir model           # would publish refuse this edit?
hashictl diff --dir model           # what changed against the Draft
hashictl propose --dir model --title "Add peak_order_share"

validate prints one line per problem: severity, file and line, a stable code, and a message. It exits with 1 when there is an error, so an agent can loop until it is clean.

test sends your folder to the server, which checks it the way a publish would. It reports:

• model tests that fail, with the rows that differ;
• evals that passed on the published model and fail now (hosted plans; the community edition skips evals);
• members you removed that stored dashboards or threads still use.

It writes nothing and exits with 1 when a publish would refuse the change.

hashictl query --json '{...}' runs a governed query as you, so the agent can check a number.

Move in from dbt and other tools

hashictl import --format metricflow --file semantic_models.yml
hashictl bootstrap --dbt target/manifest.json

import also reads ossie, snowflake, databricks and cube. Both commands open a Proposal and print what did not carry over. After an editor approves it, the agent pulls the model and maps the rest by hand.

Publish

Publishing is a person’s decision. hashictl publish --yes works only for an owner or editor, and it refuses a Draft that does not compile.

The skills

Skill	Use it to
hashi-add-metric	add a measure or a governed metric
hashi-add-dimension	add a column, a calculated field or a time grain
hashi-add-rollup	pre-aggregate a hot query
hashi-fix-model	fix validate errors, a failing model test, a lost eval or a removed member
hashi-write-evals	add agent evals to model.yaml
hashi-import-dbt	move a dbt semantic layer into Hashi
hashi-debug-query	find why a query returns the wrong rows or is refused

In Claude Code they show as /hashi:hashi-add-metric and so on, but you rarely call them by name. Ask:

Add a metric for the share of revenue from the largest order, and open a Proposal.

The agent picks the skill. It pulls the model, edits it, validates and tests it, and proposes it.