AI WorkbenchA self-hosted retrieval control plane for Astra.

What you get ​

AI Workbench is the control plane between your team's data sources and the agents you want to ground on them. It's the surface where someone on the operations side picks "ingest this PDF into the policy KB" and the surface where someone on the application side decides "this agent should retrieve from policy + tickets but not invoices."

The implementation is intentionally boring underneath: a Hono HTTP runtime, Zod at every boundary, the Astra Data API SDK for storage, and a conformance harness so the contract holds across runtimes.

Quickstart ​

Boot with the in-memory control plane — no external services required.

git clone https://github.com/datastax/ai-workbench
cd ai-workbench
npm ci && npm run install:ts

npm run dev                            # http://localhost:8080
curl http://localhost:8080/healthz     # {"status":"ok"}
curl http://localhost:8080/docs        # Scalar-rendered API reference

Want a durable control plane? Switch to Astra tables — it's a YAML change. See Configuration. The runtime can even auto-detect Astra credentials from the astra CLI if you have a profile configured.

Map the docs ​

• Product overview — workflows, mental model, and the quickstart path.
• Architecture — components, data flow, and how the layers fit together.
• Green boxes — the multi-runtime story: TS today, Python and Java scaffolds tracking the same contract.
• API spec — every /api/v1/* route with request / response shape.
• Authentication — middleware, OIDC login, silent refresh, threat model.
• Audit logging — what gets logged, the envelope, and the doc-drift test.
• Configuration — workbench.yaml schema reference.
• Workspaces — scoping, cascade semantics, defaults.
• Playground — UX for text / vector / hybrid / rerank, ingest dialog.
• Agents — personas, RAG defaults, per-agent LLM binding, conversation routes.
• MCP server — expose a workspace as MCP tools for external agents.
• Conformance — cross-runtime contract testing.
• Roadmap — phased delivery plan and open questions.

License ​

AI Workbench is MIT-licensed — fork it, embed it, ship it. Issues and pull requests welcome on GitHub.