Getting started

Install

uv tool install elbi

This puts the elbi and elbi-app commands on your PATH. (For library-only use in an existing project, uv add elbi. For just the CLI and MCP, with none of the chat app's heavier dependencies, uv tool install elbi-cli on its own.)

Later, elbi update tells you whether a newer release exists and prints the command that upgrades your particular install; see Upgrading. It only reaches the network when you run it.

Scaffold a project

elbi init acme-data-context
cd acme-data-context

init runs fully offline: no login, no cloud call. It writes:

acme-data-context/
├── elbi.yaml          # committed, logical wiring
├── elbi.dev.yaml      # local-only data bindings (gitignored)
├── derivations/churn_risk.py  # an example derivation
├── fixtures/sales.csv         # sample data for local dev + tests
├── tests/test_derivations.py  # tests of the compute functions
└── pyproject.toml

Use --template minimal for a bare project with no derivations/.

Serve it

elbi serve

No API key needed to start: this opens your browser to a chat UI at http://localhost:7700 automatically once ready, and serves the same derivations over MCP at http://localhost:7700/mcp -- point any MCP client at that URL too, and you are driving the same agent loop you would have in production. The chat UI needs a model to actually talk to, though -- add one in Settings the first time you send a message, or skip that step by exporting a key beforehand, e.g. for the default model:

export ANTHROPIC_API_KEY=sk-ant-...

The first run takes ~15-20s longer than later ones while it warms up (loading the embedding model, initializing the local database); that is expected, not a hang -- the browser opens the moment it is actually ready, not before. Pass --no-browser to skip opening one (e.g. on a headless machine).

Only want the MCP endpoint? Use elbi mcp instead -- no chat app, none of its heavier dependencies, and no LLM key of its own to configure: whatever MCP client you connect brings its own model and credentials.

elbi mcp

→ discovered 1 derivation(s) in derivations/
→ resolved data from elbi.dev.yaml
→ MCP server ready at  http://localhost:7878/mcp

Validate and test

elbi validate   # check the project against the spec
uv sync                 # only needed for uv run; serve/mcp use their own environment
uv run pytest           # run your derivation tests