Skip to main content
The content layer in Git is a Beta feature for GitHub repositories holding a dbt project. It is not available in Upsolve Local.
Most teams that run dbt already treat their repository as the source of truth for what their data means: which models exist, what each column holds, and how a KPI is calculated. The content layer lets an Upsolve workspace take everything its agent knows from that same repository, so your analytics agent and your dbt project can never drift apart. Instead of editing the data model, prompts, skills and knowledge base in Upsolve, you change them in Git:
  • A pull request is checked: the repository’s Upsolve workflow runs dbt parse on it, so a broken project fails before it is merged.
  • A merge into the branch the workspace listens to becomes a new workspace version, automatically.
  • You promote that version to production when you are happy with it, or let each merge promote itself.
Everything else — the Playground, tests, canvases, sharing, embedding — works exactly as in any other workspace.

What the repository owns

Once a workspace is connected, these parts of it come from the repository and are read-only in Upsolve. In the workspace’s rail they are grouped under one entry, Repository, with a tab for each.
A git-synced workspace's Repository entry open on Data Model: a banner says it is defined in the repository and read-only, with tabs for Data Model, Knowledge Base, Metrics, Skills and System Prompts, and the dbt models listed as tables
Still edited in Upsolve: tests, golden assets, working memory, the theme, canvases, sharing, environments and everything under Settings. These describe how the workspace is used, not what the data means, so they stay out of the repository. Metrics are owned by the repository only once a sync has found MetricFlow metrics in it. Until then the Metrics page works as in any other workspace. The full file layout, with examples, is in the Repository reference.

How a change reaches the agent

1

Open a pull request

Change a dbt model, a description, a metric, a skill or a prompt on a branch and open a pull request into the branch the workspace listens to. The Upsolve sync workflow (added by the setup pull request) runs dbt parse on it as a check. Pull request runs never change the workspace.
2

Merge it

When the pull request merges, the workflow runs again on the push and uploads its parse. GitHub tells Upsolve the run finished, and Upsolve syncs: it reads the parse, writes the workspace’s draft and publishes it as a new version.
3

Promote it

By default the new version is not what your users get yet. Check it in the Playground or run your tests against it, then promote it from the version picker. Turn on Promote each merge to production under Settings → Git to skip this step.
The sync chip at the top right of every workspace page shows the commit the workspace was last synced from and when. Click it for the repository link and Sync now, which syncs the newest successful run on the branch without waiting for a merge.
The sync chip opened: the repository and branch, the last synced commit with its table and metric counts and dbt target, a Git settings link and a Sync now button
A sync that finds nothing new — the same commit as last time — makes no version, so syncing twice is harmless.

Changing something from Upsolve

Because the repository owns these sections, Upsolve never edits them in place: a write from the editor, the API, the CLI, an MCP tool or an agent is refused. Changing one is a commit. You don’t have to leave Upsolve to make that commit. On any repository-owned page, Ask AI opens the Builder Agent with a request to propose a change. It writes the change as a commit on a branch named upsolve/changes-<branch> and opens one pull request into the branch the workspace listens to. Further changes you ask for collect on the same pull request until it is merged or closed, so an afternoon of edits is one review, not ten. The Builder Agent can propose changes to:
  • a skill, a knowledge base file or a prompt,
  • row-level security: a global rule or a model’s own condition (Team plan),
  • a model’s or column’s description and its meta.upsolve settings.
It does not write MetricFlow YAML: change metrics in the repository yourself.

Row-level security is reviewed twice

Row-level security in upsolve.yml needs the Team plan. A repository that declares rls on a lower plan is refused rather than synced without it, because dropping the rules silently would hand every user every row they were written to withhold. If a commit loosens row-level security — removes a rule, or a condition that used to filter a table — the sync stops before publishing and the workspace records what changed. The sync chip then shows the sync as failed: an owner or admin opens it, reviews what the commit loosens and accepts it, which syncs again. Otherwise, revert the commit in the repository.

Staging and production from one repository

A workspace listens to exactly one branch. To try changes before production, create two workspaces on the same repository: one listening to staging, one to main. Merging staging into main then moves the change from the first workspace to the second. Each has its own versions, tests and users. You can switch a workspace’s branch under Settings → Git, but that changes what its next sync reads; a second workspace is the safer path.

Who can do what

Disconnecting (Settings → Git) forgets the repository and stops syncing. The workspace stays a git workspace and goes back to the start of its setup page, ready to connect a repository again. The Upsolve GitHub App stays installed on your GitHub account; uninstall it on GitHub if nothing else uses it.

Next steps

Set up

Connect a dbt repository and run the first sync.

Repository reference

upsolve.yml, the upsolve/ folder and meta.upsolve.

dbt metrics

Answer from your MetricFlow metrics.