Consort Engineering Orchestrator
Keep AI-built application code on spec with deterministic workflows, real database branches, tests, and human approval gates.
- Source repo
- databricks-solutions/consort
- Stars
- ★ 33
- Last updated
- today
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 59/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude CodeCodex (Partial support)
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- A team building a Lakebase-backed transactional API can use real Postgres branches for test-driven development.
- Not a fit if
- Teams building transactional apps without Databricks Lakebase/Postgres
- Teams needing ETL, analytics, BI, or Delta Lakehouse workflows
- Source review
- 59/100 · Major gaps
What does this agent do, and when should you use it?
Consort orchestrates development of transactional application backends on Databricks Lakebase, a Postgres-compatible database, with an optional web UI. Its `consort-drive` state machine coordinates product owner, specification author, architecture reviewer, DBA, test strategist, UX designer, and Navigator/Driver roles through design, build, deployment, and promotion. Project code is paired with Lakebase branches, so tests and migrations run against real database branches. The primary workflow ships as a Claude Code plugin, with skill installation paths for other coding agents and an MCP server. It requires a Lakebase-enabled Databricks workspace and the documented development tools; it is not intended for ETL, analytics, BI, or Delta Lakehouse work.
After /consort:start, Consort can create a Lakebase-paired project or resume a project with existing .consort/ workflow state. consort-drive advances through /plan -> /design -> /build -> /deploy: role agents consume earlier artifacts and produce a structured specification, acceptance criteria, architecture and database migration plan, UX design when needed, and an ordered test list. A design gate freezes the specification. During build, the Navigator writes a failing test and the Driver implements the smallest passing change and refactors; verification runs against a real Lakebase branch, with bounded repair that does not edit the tests when verification fails. The orchestrator also handles deployment, verification, PR/CI, merge, and promotion to a parent tier, pausing for human approval at gates. The repository includes lakebase-create-project, lakebase-adopt-consort, lakebase-feature-status, commands and agent refresh CLIs, plus lakebase-mcp-server for MCP clients.
- A team building a Lakebase-backed transactional API can use real Postgres branches for test-driven development.
- A product team that wants specifications, acceptance criteria, test lists, and migration plans fixed before implementation can use the design gate.
- An engineering team that wants separate roles to write tests and review implementation can use the Navigator/Driver workflow.
- A team requiring human approval for deployment and promotion to a parent tier can use the orchestrator's gated release flow.
- Developers using a VS Code-compatible IDE, Claude Code, or an MCP client can choose the documented interface or integration path.
How do you install or deploy this agent?
Prepare a Databricks workspace with Lakebase enabled, plus Databricks CLI v1.0.0+, Node.js 20+, Python 3.10+, JDK 17+, GitHub CLI gh, and npm. The quickstart script runs the doctor, installs the Claude Code plugin, and pre-downloads the toolkit; it requires network access. Then start a project from an agent session:
How do you use this agent?
Run the start command in a new folder to create a Lakebase-paired project; a project with .consort/ resumes from its saved workflow state. Review the specification, acceptance criteria, test list, and DBA plan at the design gate, then approve deployment and promotion gates. After project creation, use the following command for CLI help:
What are this agent's strengths and limitations?
- Tests run against real, isolated Lakebase Postgres branches rather than database mocks that may drift from production.
- A state machine routes phases, and fail-closed gates prevent agents from advancing without human approval.
- Navigator and Driver separate test authorship from implementation, and tests cannot be changed to force a pass within a unit of work.
- Alongside the Claude Code plugin, the repository documents skill installation, an MCP server, and a pre-rendered OpenAI Foundry tool spec.
- The core workflow depends on Databricks Lakebase and a workspace with the feature enabled; it is not a generic database or local-only test orchestrator.
- Initial setup requires several runtimes, the Databricks CLI, and workspace authentication; the docs describe this as infrastructure-heavy setup.
- The primary plugin workflow is tied to Claude Code; other agents use distinct skill or MCP paths and may have different integration capabilities.
- The repository does not state software or Databricks/Lakebase service pricing, so total cost needs separate confirmation.
How does this agent compare with similar options?
The README compares Consort with GitHub Spec Kit and superpowers. Consort enforces its process with a deterministic state machine the agent cannot edit, fail-closed human approval gates, immutable tests, and runs against real Lakebase branches. Spec Kit is characterized as spec-first followed by a trusted build, while superpowers asks the model to follow test-first rules. Consort targets transactional Lakebase/Postgres applications; the README describes the other two as general software workflows.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Consort Engineering Orchestrator This agent | 59 · Major gaps | CLI | ★ 33 | today | TypeScript | Claude Code |
| Pilot Shell | 53 · Major gaps | CLIPaid | ★ 2.1k | 4d ago | JavaScript | Claude Code |
| Agent Skills — Production-Grade Engineering Skills for AI Agents | 71 · Some gaps | CLIFree + model costs | ★ 106k | 2d ago | JavaScript | Codex · Claude Code · OpenAI API |
| ECC Engineering Harness | 66 · Some gaps | Agent plugin / skillFreemium | ★ 278k | today | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README says deployment, upgrades, migrations, and tests access project code, GitHub, Databricks/Lakebase, and npm, and claims telemetry sends only allowlisted fields with CONSORT_TELEMETRY=0 to disable it. The install command executes a downloaded curl script, so operators must assess that trust chain. Human approval gates and fail-closed behavior are described, but the supplied evidence does not verify that every side effect is gated or recoverable. SECURITY.md provides a vulnerability reporting email; dependencies are listed without lock, audit, or mitigation evidence here. Unverified publisher identity is not a deduction; repository attribution and a reporting channel are present.
The documentation consistently describes phases, gates, live Lakebase branches, and prerequisites; the CI workflow shows typecheck and test jobs. Availability depends on external CLIs, services, databases, and npm. For failure messages, the README says the doctor checks prerequisites and gives fix hints, but the supplied material does not cover the full range of failures. Tests were not executed for this static review.
The audience and use case are well bounded: coding agents building transactional backends on Lakebase; ETL, analytics, Spark, and Delta Lakehouse use are explicitly excluded. Several editor and agent entry points are listed, but the workflow requires Lakebase, a Databricks workspace, and a specific toolchain; triggers are described mainly through /consort:start and project state.
The README is organized across setup, workflow, installation, updates, repository contents, CLIs, support, and contribution, and points to a StockFlow example and positioning/FAQ material. Prerequisites and update operations are detailed, though users still need to evaluate the one-command curl installer and differences between install paths. Names are mostly stable, with numerous aliases adding some complexity. LICENSE.md contains a restrictive DB license, which is not equivalent to a permissive open-source license despite the NOASSERTION metadata. A package version and release/update flow are described, but no changelog is supplied. Contribution and issue channels identify maintenance paths, while named responsibility remains unclear.
The product offers role separation, a state machine, approval gates, real database branches, and project-state/CLI outputs, with clear value for its narrow use case. The supplied material does not independently establish outcomes, and the Lakebase/Databricks, Claude, or other tooling requirements create setup and service costs, so cost-benefit receives partial credit.
README claims about setup, dependencies, telemetry, workflow, and boundaries have supporting package, CI, license, and security-policy material. Telemetry details are summarized only in the README; its referenced implementation and TELEMETRY.md are not supplied. Cross-source corroboration is limited, mainly between README descriptions and configuration files. The rationale distinguishes documented claims from directly visible file evidence and does not treat claims as executed results.
- The install command executes a script fetched from the network; deployment and tests also access Databricks/Lakebase, GitHub, and npm. Review the script and its permissions, and configure credentials in a controlled environment.
- The telemetry contract and opt-out are summarized in the README, but TELEMETRY.md and implementation files were not supplied, so their data boundaries could not be checked statically.
- The license restricts use of the materials to use with Databricks Services. Review its full terms before adoption, redistribution, or dependency.