databricks-solutions/consort

Consort takes its name from the field of music. A *consort* is an ensemble that plays in concert: each musician holds one part, and a conductor keeps them in time. Consort is that, applied to building software. A set of agents each take on one familiar role from the software lifecycle — a spec author, an architect, a DBA, a test strategist, a UX designer, and a navigator/driver pair at the keyboard — while the product owner is the human's own seat (you, or a Human Proxy standing in headless), a deterministic conductor keeps them in sequence, and the human approves every gate. No agent plays another's part.
**What Consort builds.** Consort builds **transactional application backends** (an API, and optionally a web UI) on Lakebase, Databricks' serverless, Postgres-compatible OLTP database that branches in about a second. Every git branch is paired with its own Lakebase branch, so each has a real, isolated database and the schema evolves in lockstep with the code. It isn't for the Delta Lakehouse, ETL, analytics, or BI; it's for building app backends on Postgres. See docs/positioning.md for how Consort compares to other spec-first tools.
Quick start
Setting up Consort is easy. Copy the command below and enter it into any terminal window.
bash
bash .)
**What to expect.** Consort drives the loop /plan -> /design -> /build -> /deploy and stops at every gate for you:
- at the **design gate**, you review and approve the frozen spec: the stories and acceptance criteria, the ordered test list, and the DBA's schema plan;
- through the **build**, each cycle writes a failing test, makes it pass against a live database branch, then refactors;
- at the **deploy** and **promote** gates, you approve the release and the migration to the parent tier.
Nothing advances past a gate without you.
**Walk through a full first project.** examples/first-project/ is a step-by-step walkthrough of one session, install to first shipped feature, using a sample inventory app (StockFlow). It ships copy-ready seed files so you can launch your own project from ours instead of starting from a blank page.
What it installs, runs, and sends
- **Install-time:** the bootstrap downloads the Consort toolkit from this repository, and the plugin's SessionStart hook runs npm ci in the plugin directory to finish installing its dependencies (packages from the npm registry).
- **Runtime:** Consort spawns role agents as claude --agent processes inside your project, and applies schema migrations and runs tests against your Lakebase database over your Databricks workspace connection. It runs only inside the project you point it at.
- **Telemetry:** pseudonymous usage telemetry (Level 1 by default: one trace per run, drawn from a closed allowlist), plus a one-time install beacon carrying only a random install id, the kit version, and a timestamp. No free-form data, code, prompts, or project content is ever sent. Level 2 is a separate opt-in. The full contract, including how to turn it off (CONSORT_TELEMETRY=0), is in TELEMETRY.md.
Other ways to install
For coding agents other than the Claude Code plugin, install.sh copies the skill trees under skills/ into the path each agent reads from, pulling the latest vendored skills first (best-effort; skipped offline). It auto-detects installed agents; --tools overrides.
bash
claude plugin marketplace update databricks-solutions
claude plugin update consort@databricks-solutions