Falk

Governed AI access to your data warehouse, powered by semantic layers.

Define metrics once in YAML. Query them naturally through Slack, CLI chat, or MCP. Same numbers everywhere.

Why Falk

Most AI SQL agents can answer fast but drift on metric definitions. Falk puts the semantic layer in charge, so AI uses approved metrics and dimensions only.

• One metric definition in YAML, reused everywhere
• Governed access instead of free-form warehouse querying
• Consistent numbers across Slack, CLI, web, and MCP

What you get

• Governed access — only approved metrics and dimensions, not raw tables
• MCP server — works with Cursor, Claude Desktop, and any MCP client
• Consistent numbers — same as your BI layer
• Business context — synonyms, gotchas, and rules built into the schema
• Multi-interface — MCP server, Slack, CLI chat — same data everywhere

Quick start

Get a working project in under 5 minutes:

git clone https://github.com/Fredehagelund92/Falk.git
cd Falk && uv venv && uv sync

falk init my-project
cd my-project

cp .env.example .env   # Add your OPENAI_API_KEY
falk validate --fast

# Start querying
falk chat      # Web UI at http://127.0.0.1:8000
# or
falk mcp       # Connect from Cursor or Claude Desktop
# or
falk slack     # Deploy for your team

Learn more

Topic	Description
Quick Start	Full setup walkthrough
State & Memory	What persists across requests and sessions
Learning & Feedback	How the agent improves over time
Semantic Models	Define your metrics
CLI Reference	All commands

Deployment note (0.1.0)

• Alpha — not battle-tested for production yet. Use at your own risk.
• Recommended: one company/workspace per deployment (single-tenant).
• In production (FALK_ENV=production), configure access_policies to avoid open access.
• Session storage defaults to memory. For production, set session.store: postgres and POSTGRES_URL in .env. See Memory.
• Chart generation uses session state; rerun the query if a restart drops chart context.
• Slack exports default to DM-only; optionally allow channels via slack.export_channel_allowlist.

Inspiration & Credits

falk was inspired by excellent work in the data agent space:

• OpenAI’s in-house data agent — Grounded metric querying and data agents
• nao — Context engineering patterns and agent reliability testing
• dash — Self-learning from feedback and six layers of context

We’re grateful to these projects for showing what’s possible with well-designed data agents.

This site is open source. Improve this page.