Open source · Apache-2.0
Elbi is an open-source, AI-native data science environment that runs on your own machine. It’s built on Intelligible’s core principle: answers should come from components you can inspect, not from what a model remembers.
In Elbi, those components come from derivations, and everything else is built on them. A derivation is a Python function with declared inputs and a declared output. Elbi serves each one over the Model Context Protocol as a versioned, cached tool. It runs in your own process, and there’s no account to sign up for.
Verification gates
33 checks an agent can call over MCP, covering regressions, survival models and forecasts as well as leakage, fairness and difference-in-differences.
Agent-authored derivations
An agent can propose a derivation. Elbi runs it in a sandbox and serves it once it passes verification, or holds it for a person to approve if you turn on manual certification.
Signed certificates
Ed25519-signed and checkable offline. Each one records what was checked and who issued it.
Chat with verdicts
The web app opens on a chat. Every answer carries a verdict from the gates and a certificate to download.
A warehouse
Connectors for SQL databases, MongoDB, DynamoDB, Salesforce, HubSpot, Stripe, Shopify, Zendesk, Google Sheets and object storage.
Models and orchestration
Model training, scheduled runs and monitors, all in the web app.
Result history
See how a certified answer changed between runs, and whether the data, the code or the parameters moved it. Versions can go to MLflow.
SQL workbench and notebook
Explore without any gates, then promote a query or a notebook cell to a derivation.
Dashboards, feature store, metrics
Dashboard tiles bound to derivations, point-in-time-correct training joins, and a metrics layer that imports and exports OSI.
Monitoring and lineage
Alerts when a verified number moves, with the verdict attached. Provenance and impact analysis across every artifact.
Data contracts check a cleaning step for types, ranges, keys and referential integrity before it’s certified. Profiling and contract checks can push down to DuckDB or Polars when the data doesn’t fit in memory.
Elbi runs on your machine, on Python 3.10 or newer. There’s no account and no sign-in, and nothing leaves your network unless you send it somewhere.
There are two ways to work with Elbi. Connect it to your AI agent over MCP, and the agent can call your derivations and Elbi’s verification gates as tools, using whatever model it already runs. Or open the web app and ask about your data in the chat: each answer shows a verdict (sound, inconclusive or unverified) with a certificate you can download.
$ uv tool install elbi$ elbi init acme-data-context$ cd acme-data-context$ elbi serveYou don’t need an API key to start. A chat UI comes up on localhost:7700 and the same derivations serve over MCP at /mcp. The first run takes 15 to 20 seconds longer than later ones, because it downloads an embedding model and sets up the local database.
If you only want the endpoint, install elbi-cli and run elbi mcp. That leaves out the chat UI and the heavier dependencies. There’s no model key to set either: Elbi doesn’t ship a model, and the client you connect brings its own.
Guides and reference live in the repo’s docs, next to the code.
Open-source Elbi is single-user. It keeps one person’s derivations, cache and certificates on one machine.
Enterprise puts a single Elbi deployment in front of the team, with sign-in, access control and tenant isolation. Everyone works from the same registry of derivations and the same certificate history. It deploys with Terraform to AWS, Azure or GCP.
Commercial use included.