Baton

A metadata-grounded codegen relay for DataHub.

🔗 Live demo: baton.endpx.cloud — no signup, no setup.

Open the Studio and press Run. It hits a real DataHub instance (827 entities from the showcase-ecommerce datapack): the Context agent searches the catalog, stops to ask which orders you meant, pulls the real columns, and the run ends with a validated dbt model and a generated-by-baton tag written back onto the source datasets. ▶ Demo replays a recorded trace if you would rather not wait for a live run.

The app is co-located with DataHub on its own host, because GMS is bound to localhost — a cloud-hosted frontend would have no route to the catalog it claims to read.

Baton is a multi-agent pipeline that turns a natural-language goal — "generate a dbt model joining orders and customers, filtered to the last 90 days" — into a validated, PR-ready dbt model, grounded in what your DataHub catalog actually knows: real schemas, real lineage, real ownership.

Like a relay race, three specialized agents each run their leg and pass a structured baton of context forward — and the last runner writes the result back into DataHub, so the next person (or agent) inherits the knowledge.

🏆 Built for Build with DataHub: The Agent Hackathon — Track 1: Agents That Do Real Work.

How it works

  goal ──▶ [ Context agent ] ──▶ [ Codegen agent ] ──▶ [ Publisher agent ] ──▶ dbt model + write-back
             resolves entities      generates SQL          packages .sql/.yml
             pulls schema &         validates against      writes provenance
             lineage via MCP        real schema (sqlglot)  back to DataHub
                                    self-corrects

1. Context agent — resolves the tables you mention into DataHub URNs via the DataHub MCP Server (search, get_entities), then pulls real column names/types and lineage (list_schema_fields, get_lineage, get_lineage_paths_between).
2. Codegen agent — generates a dbt SQL model constrained to the fetched schema, then validates every column reference with schema-aware sqlglot qualification. Validation errors feed a bounded self-correction loop.
3. Publisher agent — packages the validated SQL into a dbt .sql model + .yml schema file, and writes provenance back to DataHub (tags/descriptions on the source tables), closing the loop in the metadata graph.

Every MCP tool call streams to a live trace UI, so you watch the baton move between agents in real time.

Status

🚧 Under active development for the hackathon (deadline Aug 10, 2026).

Self-hosting

Requirements: a DataHub instance (e.g. datahub docker quickstart) and an API key for any OpenAI-compatible LLM endpoint — the hosted demo runs meta/llama-3.1-70b-instruct on NVIDIA NIM, but OpenRouter, Together or a local vLLM work by changing LLM_BASE_URL.

# 1. DataHub, with the demo catalog
pip install acryl-datahub
datahub docker quickstart
datahub datapack load showcase-ecommerce      # ~827 entities

# 2. The SQL validator (sqlglot lives in Python, the orchestrator does not)
cd validator && python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/uvicorn app:app --host 127.0.0.1 --port 8100 &

# 3. Baton
cd ../web && npm ci
cp ../.env.example .env.local                 # fill in LLM_API_KEY + DATAHUB_GMS_TOKEN
npm run build && npm run start -- -p 3400

uvx must be on PATH — Baton spawns uvx mcp-server-datahub@latest as its MCP sidecar, because GMS has no built-in /mcp endpoint on the open-source build. Get the DataHub token from Settings → Access Tokens.

Run it on the same host as DataHub. GMS should stay bound to localhost, which means a frontend hosted anywhere else cannot reach it.

Sample output

examples/ holds a dbt model Baton actually generated, with the run that produced it written down — including why the column names in it are the proof that grounding works.

Tech

Next.js (TypeScript) · DataHub MCP Server · sqlglot · any OpenAI-compatible LLM · XyFlow (react-flow)

Credits and disclosures

Baton was built during the hackathon submission period. Beyond the usual frameworks and libraries, these pieces came from elsewhere:

• shadcn/ui (MIT) — the badge, button and card primitives in web/src/components/ui/, copied as that project intends.
• Radial orbital timeline — the animation in the landing page's The relay section began as a community-shared React component. It was restructured for Baton: React 19 ref semantics, a container-measured orbit radius, and the node data replaced with the pipeline's real stages and the tools each one calls.
• DataHub (Apache 2.0), sqlglot (MIT), Next.js, XyFlow, Tailwind CSS, lucide — used as dependencies.

Everything else — the three agent lanes, the graph rules, the stage palette, the validation loop, the write-back, the trace UI — is original to this submission.

License

Apache 2.0