DataFoundry is an out-of-the-box, self-hostable enterprise data agent workbench that upgrades natural-language data analysis into semantically aware, secure, and verifiable data workflows, helping teams reach trusted answers faster while preserving audit-grade analytical evidence.
English · 简体中文
Quick Start · Docs · Supported Data Sources · Contributing · License
Modern data agents need more than a chat model. They need selected context, datasource boundaries, SQL policy, auditable events, durable outputs, and a frontend protocol that can replay the whole run.
DataFoundry puts those pieces behind one runtime:
- 🔎 Schema-first analysis — the agent inspects datasource structure before it can run read-only SQL.
- 🧠 Governed context — conversation history, memory, tool results, files, and knowledge sources are compiled under one budget.
- 🧾 Auditable execution — AG-UI events, SQL audit logs, artifacts, and session history are persisted as replayable records.
- 📦 Unified assets — uploads, workspace files, generated outputs, and KB imports share the same deduplicated asset layer.
- 🧩 Protocol-ready runtime — CopilotKit / AG-UI clients consume the same events, run state, artifacts, and replay data.
DataFoundry is built around a Data Gateway adapter boundary. The current runtime already recognizes local files, embedded databases, cloud warehouses, lakehouse engines, operational databases, and search / NoSQL systems.
The frontend talks to a single backend runtime. The backend owns identity, run replay, context assembly, memory, tool policy, SQL guardrails, file references, and artifact creation. The model sees a governed prompt; it never sees raw datasource credentials. 🛡️
The recording below shows the current web workbench experience.
The recording below shows the current terminal workflow backed by the same runtime.
npm install
cp .env.example .env
cp apps/web/.env.example apps/web/.env.local
npm run devOpen the workbench:
http://127.0.0.1:3000/data-tasks
The local workbench includes a demo DuckDB datasource. Live agent runs require a real LLM key in .env.
LLM_PROVIDER=openai-compatible
LLM_MODEL=qwen-plus
LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
LLM_API_KEY=replace-with-your-key
DeepSeek and other OpenAI-compatible providers use the same provider mode:
LLM_PROVIDER=openai-compatible
LLM_MODEL=deepseek-chat
LLM_BASE_URL=https://api.deepseek.com
LLM_API_KEY=replace-with-your-key
| Use case | Runtime support |
|---|---|
| Natural-language database analysis | Datasource selection, schema inspection, SQL guard, query limit, timeout, audit log, table artifact. |
| File-backed agent work | Session workspace, cross-session workspace assets, file refs, downloads, generated deliverables. |
| Knowledge-assisted analysis | KB imports, document chunks, local search, optional embedding-backed retrieval, governed context injection. |
| Frontend agent UX | CopilotKit / AG-UI streaming, run replay, task state, token usage, artifacts, interaction suspension. |
| Controlled tool extension | Mastra tools, MCP middleware, workspace tools, skill packages, tool observation adapters. |
npm run build
npm run smoke:config-api
npm run smoke:data-gateway
npm run smoke:copilotkit
npm run smoke:docsUse targeted smoke checks for the package you touch. package.json lists the full verification set.
DataFoundry is moving quickly, so small, well-scoped contributions are easiest to review.
See CONTRIBUTING.md for setup, verification, and hygiene rules.
- Open an issue or discussion for behavioral changes, protocol changes, datasource adapters, and agent-policy changes.
- Keep pull requests focused on one runtime boundary or feature area.
- Run
npm run buildand the targeted smoke checks for the packages you touched. - Update docs when a change affects setup, APIs, datasource configuration, event behavior, or user-visible output.
- Do not commit credentials, local databases, generated storage, or private benchmark data.
- Semantic data operating layer - Build a durable business semantic layer for metrics, entities, joins, lineage, policies, and reusable analytical concepts.
- Autonomous analyst loops - Let agents plan investigations, run controlled experiments, critique findings, and converge on evidence-backed conclusions.
- Evaluation and reliability lab - Create repeatable NL2SQL, retrieval, tool-use, and end-to-end task benchmarks with regression gates and failure forensics.
- Multimodal knowledge fabric - Unify tables, documents, notebooks, charts, images, logs, and generated files into one governed context and retrieval fabric.
- Agent app platform - Expose DataFoundry as a platform for domain-specific analytical agents, reusable workflows, custom tools, and shareable agent apps.
- Enterprise control plane - Add multi-tenant governance for identity, RBAC, approvals, audit export, policy-as-code, cost limits, and deployment operations.
- 2026-07-01: DataFoundry first complete release - DataFoundry now ships as a full-stack governed AI data workbench: a Web workbench and TUI share one TypeScript agent runtime, CopilotKit / AG-UI event stream, replayable run history, task state, SQL audit trail, artifact output, and unified file asset layer. The first release connects natural-language analysis to real datasource registration, connection testing, schema introspection, table preview, read-only SQL execution, row limits, masking, knowledge imports, workspace files, MCP resources, packaged skills, a built-in data-analysis skill, and model configuration. In short, DataFoundry has moved from an agent demo into an extensible data-agent operating foundation for safe, traceable, and reusable analytical work.
| Quick Start Install, configure a model key, and run the workbench. |
Product Overview Understand the product positioning and analysis workflow. |
| Supported Data Sources Datasource types, fields, and connection boundaries. |
Agent Runtime CopilotKit / AG-UI run input, events, and safety boundaries. |
| REST API HTTP endpoints for local development and integration. |
Architecture High-level runtime, Data Gateway, files, knowledge, and artifacts. |
DataFoundry is under active development. Current code, public docs, and passing smoke checks are the source of truth.
Apache License 2.0. See LICENSE.