Build & deploy Databricks AI agents in minutes

BrickForge is a pip-installable tool that takes you from zero to a live, production-grade AI agent on Databricks Apps. No code, no notebooks, no YAML wrangling.

Install from PyPI View on GitHub

$ pip install brickforge && $ brickforge

Beta

How It Works

Five steps. One tool. Zero boilerplate.

A visual setup app walks you through everything. Each step is a block in a directed graph - configure, execute, done.

Connect

Authenticate to any Databricks workspace via one-click OAuth
Token stored in your OS keychain, never on disk
Pick a SQL warehouse and Unity Catalog schema

Generate

Describe your domain in plain English
AI generates table schemas, synthetic data, SQL functions, stored procedures
System prompts and knowledge base tailored to your use case

Wire

Pick a Foundation Model endpoint
Connect Genie, KA, Vector Search, APIs, MCP, A2A agents
Toggle features like memory and charts

Deploy

One click. Bundles code + config + chat UI, deploys as a Databricks App
Auto-grants every permission the app needs: tables, functions, warehouse, Genie, serving endpoints, Lakebase
No manual SP configuration, no chasing permission errors. You get a live URL

Iterate

Save projects, export as .forge bundles, share with colleagues
Push to GitHub. Switch between workspaces
Clean up resources when done

Setup Panel

18 blocks. Every resource covered.

Each block follows the same pattern: choose an approach, configure, execute, done. No terminal needed.

1 Workspace 2 SQL Warehouse 3 Unity Catalog 4 Data Tables 5 Functions 6 Model Endpoint 7 Agent Prompt 8 Genie Space 9 Agent Bricks 10 Vector Search 11 MCP Servers 12 REST APIs 13 A2A Agents 14 Features 15 Lakebase 16 MLflow 17 Deploy 18 GitHub

Capabilities

Everything an agent needs. Built in.

BrickForge wires together the full Databricks platform into a single deployable agent. Here's what your agent ships with.

AI Data Generation

Describe a domain, get relational table schemas
Synthetic CSV data, UC functions, stored procedures
5-layer self-healing SQL pipeline

Genie NL-to-SQL

Agent asks Genie questions in natural language
Genie writes the SQL, queries your tables, returns results
No prompt engineering for SQL generation

Knowledge Assistants

Upload PDFs, DOCX, HTML, markdown
Grounded answers with source citations
Vector Search as complementary retrieval path

External APIs & MCP

Connect any REST API or MCP server as a tool
UC-governed connections or direct HTTP with auth
Each becomes a callable agent tool

One-Click Deploy + Auto Grants

Bundles agent code + chat UI + config, deploys via Databricks SDK
Detects app service principal, grants SELECT on tables, EXECUTE on functions, CAN_USE on warehouse, CAN_RUN on Genie, CAN_QUERY on endpoints
Zero manual permission work. No more chasing 403s after deploy

Per-User Memory

Long-term memory backed by Lakebase
Remembers preferences, facts, context across sessions
Save, recall, delete - scoped per user

Agent Tools

What your deployed agent can do

Every tool is auto-discovered, dynamically loaded, and feature-gated. No hardcoding.

Tool	How it works	Discovery
UC Functions	Auto-discovered from your schema. Parameterized SQL queries as callable tools.	Auto
Stored Procedures	CALL proc(params) for data mutations. Update records, trigger workflows.	Auto
Genie MCP	Natural-language questions translated to SQL via Databricks Genie.	Config
Knowledge Assistant	RAG over uploaded documents. Grounded answers with citations.	Toggle
Vector Search	Semantic document retrieval via Databricks Vector Search index.	Config
External APIs	REST calls via UC connections or direct HTTP. GET, POST, PUT, DELETE.	Config
MCP Servers	Any MCP-compatible tool server. Weather, Slack, custom services.	Config
A2A Agents	Delegate to remote agents via Google A2A protocol (JSON-RPC).	Config
Charts	Generate bar, line, area, and pie charts inline in chat.	Toggle
Memory	Per-user long-term memory. Save, recall, delete across sessions.	Toggle
Custom Tools	Drop a Python file with @tool functions into tools/. Auto-loaded.	Auto

Architecture

What gets deployed

Two services in one Databricks App. The agent + the chat UI.

Databricks App (your workspace) | +- MLflow AgentServer [port 8000] | |-- LangGraph agent (LangChain tools + Foundation Model) | |-- /invocations endpoint | +-- Tools: Genie, KA, Vector Search, UC Functions, | APIs, MCP, A2A, Memory, Charts | +- Chat UI [port 3000] |-- React frontend (Vercel AI SDK) |-- Express backend (auth, chat proxy, SSE) +-- Live data panel, structured response blocks, action buttons, inline charts Config: Single config.json shipped in bundle Agent reads at boot, flattens to env vars No app.yaml env vars needed

Portable Agents

Build once. Deploy anywhere. Own your code.

Every agent you build is fully portable. Export it, version it, share it, deploy it to any workspace, or take over the code entirely.

.forge Bundles

Your entire agent in one file. Config, data schemas, SQL functions, stored procedures, system prompt, knowledge base - all packaged as an open, portable .forge.zip bundle.

Workflows this enables:

Design an agent locally, export the bundle, deploy to a test workspace
Share the bundle with a customer, partner, or another team
Import into a production workspace - different host, different schema, same agent
Store bundles as versioned snapshots - roll back anytime
Hand off a fully working agent without sharing credentials or access

GitHub Export

Push your entire agent to a private GitHub repo in one click. All generated code is yours - versioned, readable, ready to fork.

What you get:

Full agent source: LangGraph runtime, tool definitions, SQL functions, prompts
Generated data layer: table schemas, seed CSVs, stored procedures
Config as code: config.json with all wiring (tokens stripped)
Vibe code on top - use Claude Code, Cursor, or any AI IDE to customize
Your agent becomes a real software project, not a locked platform artifact

Use Cases

Adapt to any domain

BrickForge ships with a flight-operations reference agent, but the data gen wizard lets you build for any industry.

Airport Operations

Check-in Performance Advisor

Monitor zone metrics, detect bottlenecks, redeploy staff
Live dashboard refreshes as the agent acts
Ships as the reference implementation

Financial Services

Portfolio Risk Monitor

Query positions, run compliance checks
Generate allocation reports
Reads from UC tables, answers from regulatory docs via KA

Healthcare

Clinical Trial Assistant

Track enrollment metrics, adverse events, site performance
Query structured trial data
Retrieve protocol documents in one conversation

Retail & Supply Chain

Inventory Ops Agent

Monitor stock levels, forecast demand
Trigger replenishment procedures
Connect POS data, warehouse systems, supplier APIs

Energy & IoT

Grid Monitoring Agent

Ingest sensor data, detect anomalies
Dispatch maintenance
Combine real-time Genie queries with equipment manuals

Your Domain

Describe it. Build it.

Type a domain description
BrickForge generates tables, functions, prompts, knowledge base
Review, tweak, deploy. Any domain, any use case.

Evaluation

Automate your agent's evaluation

Ship with confidence. BrickForge includes a built-in MLflow evaluation pipeline so you can measure agent quality before every deploy.

LLM Judge Scoring

Custom Claude-based LLM judge evaluates every response
Scores on correctness, relevance, groundedness, and completeness
Baseline vs with-guideline comparison in a single run

MLflow Tracking

Results logged to MLflow experiments - compare runs side by side
Track score distributions, failure modes, regressions
Ships with curated test datasets for the reference domain

Iterate on Quality

Tweak prompts, tools, or knowledge base - re-run eval to measure impact
Catch regressions before they reach production
Build your own test datasets from real conversations