portia
GitHub Install

Open source · early

It learns your data. And it doesn't forget.

A copilot for data scientists. portia understands what your data sources are, keeps what it learns, and builds the pipeline with you.

Install See it running

Data has no value without context. portia builds yours.

Inside the app

It measures, it builds, and it remembers.

Watch it work.
It reads, measures and understands every source you give it. Everything is grounded, nothing is hallucinated.
The portia workspace. The bookings source's catalog entry fills the middle: a written summary of the table, a dated note, then every column with its type, its role, how much of it is null and how many distinct values it holds. In the right pane an indexing job is reading the five sources, and its note on bookings waits for Allow or Deny.
The sources and indexing pane with five sources ticked, each marked not indexed, and a button reading "Index 5 sources".
The same pane one step later, each source marked not read and the button now reading "Interpret 5 sources".
Ask it anything. It answers in charts, drawn straight from your data.
A line chart of booked EUR room revenue by stay month, one line per booking channel, open in a tab beside the chat that asked for it.
The chat turn behind the chart: the question, then the copilot describing the source, querying it and plotting the result, with a note that it kept to EUR because the project has no exchange rates.
It builds every step of the pipeline on something measured.
The pipeline canvas after a run. Three sources flow through staging and intermediate models into one hotel-month table. Below the canvas is that table's measured outcome: its inputs, its row count, its columns and a unique grain. The copilot's summary of the build is in the right pane.
The copilot asking how cancelled bookings should affect rooms sold in the hotel-month table, marked as waiting for you, with the first two of its answers visible.
It keeps everything it learns, so nothing starts from scratch.
The knowledge graph by column, zoomed in on its middle. Each column is a node around the source or model it belongs to, with arrows tracing where it came from, and a group the copilot recorded links the property tables.

How it works

It reads your data. All of it.

Not by scrolling through it. By asking it questions, and remembering the answers.

  1. 01

    Everything is measured.

    Every fact it works from is one it went and got. Nothing is inferred from a sample.

  2. 02

    No hallucinations.

    Each step of the pipeline stands on something measured. Nothing is invented along the way.

  3. 03

    Everything is remembered.

    Everything it learns joins what it already knows. The tenth question is easier than the first.

Wherever your data is

It runs where your data already lives.

The warehouse the team administers, and the file somebody sent you on Friday. Neither one has to be moved, reshaped or loaded anywhere first.

  • Snowflake
  • BigQuery
  • Databricks soon
  • AWS soon
  • Azure soon
  • Spark soon
  • Trino soon

Or whatever is already on your machine

  • Postgres
  • DuckDB
  • ClickHouse soon
  • MySQL soon
  • SQLite soon
  • Sheets soon

Nothing gets copied. No extract, no staging table, no second version to keep in sync.

Size is not your problem. A table too big to open is one portia still knows.

Whichever model you trust

Any model. Including the one on your laptop.

The model never sees your rows, only what the measurements came back with. Which is why you do not need the biggest one.

  • Anthropic
  • OpenAI
  • Gemini soon
  • Mistral
  • Llama
  • DeepSeek
  • Qwen
  • xAI soon
  • Cohere soon
  • Groq soon
  • Bedrock soon
  • Vertex AI soon
  • OpenRouter soon
  • Together soon

Or entirely on your own hardware

  • Ollama
  • LM Studio soon
  • vLLM soon
  • Hugging Face

Bring your own. The provider you already pay for, or none at all.

Small models are enough. The context does the work, so your bill stays small too.

Install

It's early. Come and break it.

Free and open source. It runs on your machine, against your data. 


              git clone https://github.com/Jad1908/portia.gitcd portiauv sync --extra ui --extra agent --extra graphuv run python -m portia.ui

Needs Python 3.11 and uv. Read the README 


              Install portia for me by following https://github.com/Jad1908/portia/blob/main/INSTALL.md

Paste it into Claude Code, Codex or any agent that can run commands. Read INSTALL.md

Not ready to install it yet? Leave an address. You hear from portia when there is a new version, and at no other time.

A person reads these.

Questions

The obvious objections.