AI Data Science IDE Erdos Forks VS Code, Open Source
By: Evgeny Padezhnov
Most AI coding tools target software developers, not data scientists. Erdos changes that. It is an open-source IDE built specifically for people who live across Jupyter notebooks, R scripts, and SQL queries.
Erdos comes from Lotas, the team behind Rao — an AI assistant for RStudio. Rao brought a Cursor-like experience to R users. Erdos extends the same idea to the full data science stack: Python, R, SQL, and Julia. The project is a fork of VS Code, licensed under AGPLv3, with source code on GitHub.
Why a Data-Science-First IDE Exists
Data scientists are often second-class citizens in modern editors. VS Code, Cursor, and Claude Code are designed for application developers. Their workflows assume files, functions, and repositories — not notebooks, plots, and database connections.
The mismatch shows up daily. A data scientist switches between a Jupyter notebook, a standalone R script, and a SQL console. Generic AI assistants struggle here. They treat a notebook as plain text. They lose the execution context. They cannot reach the active R or Julia session.
Erdos addresses the gap by building the AI into the data science environment, not bolting it onto a code editor. Key point: the AI has access to the same consoles, variables, and data sources the human uses.
What Erdos Includes
The feature set maps directly to a data science workflow. Each pane solves a concrete task.
AI Across All File Types
The assistant can search, read, and write across Python, R, SQL, and Julia files. For Jupyter notebooks, Erdos uses an optimized jupytext system. Jupytext represents notebooks as plain scripts. This lets the AI make faster edits without parsing heavy JSON cell structures.
Common mistake: editing a .ipynb file as raw JSON. The format mixes code, output, and metadata in one blob. Diffs become unreadable and edits break the notebook. The jupytext approach sidesteps that problem.
Built-In Consoles
Erdos ships Python, R, and Julia consoles. Both the user and the AI can run code in them. The AI does not work blind. It can execute a command, read the result, and adjust. In plain terms: the assistant shares the live session instead of guessing at output.
Plot, Database, and Environment Panes
The plot pane tracks and organizes plots by file and time. No more lost figures across a long session.
The database pane connects to SQL or FTP data sources. Users manipulate data directly from the IDE. The environment pane shows variables, packages, and active environments — the standard view R and Python users expect.
A help pane renders documentation for Python, R, and Julia. These panes mirror the RStudio layout that many data scientists already know.
Remote Development
Erdos supports remote work over SSH or containers. This matters for teams that run heavy jobs on remote machines or GPUs. The IDE stays local; the compute stays remote.
How the AI Backend Works
Erdos offers three ways to power the assistant. The choice affects privacy and cost.
- Single-click sign-in to the Lotas zero data retention backend. Code sent for inference is not stored.
- Bring your own key. Plug in an API key from a provider you already pay for.
- Local model. Run inference on your own hardware. Nothing leaves the machine.
Key point: the local-model path makes Erdos viable for regulated environments. Healthcare, finance, and government teams often cannot send code to external APIs. A fully local option removes that blocker.
Bring-your-own-key gives cost control. Teams route requests through their existing provider contract instead of a new subscription.
How Erdos Compares to the Alternatives
A short comparison clarifies where Erdos fits.
| Tool | Built for | Notebook-native | Open source |
|---|---|---|---|
| VS Code | General development | Partial | Yes (MIT) |
| Cursor | Software engineers | Partial | No |
| Erdos | Data scientists | Yes (jupytext) | Yes (AGPLv3) |
| RStudio | R users | R Markdown | Yes |
Cursor is a strong AI editor but closed source and developer-oriented. RStudio is data-science-native but centered on R. Erdos aims for the overlap: native data science panes plus multi-language AI, with the source open.
The AGPLv3 license is worth attention. It is a strong copyleft license. Any networked deployment that modifies Erdos must release its source. For internal use this is rarely an issue. For vendors building a hosted product on top, it is a real constraint. Read the license terms before commercial redistribution.
Practical Notes for Evaluation
Erdos is a fork of VS Code. Existing VS Code muscle memory and many extensions should carry over. That lowers the switching cost for anyone already in that ecosystem.
The project is new and community-driven. The team explicitly asks for feature requests. Early adopters can shape the roadmap. The trade-off is maturity — expect rough edges that a years-old IDE would not have. Test it on a non-critical project first.
For Jupyter-heavy teams, the jupytext editing path is the feature to evaluate first. Measure whether AI edits to notebooks stay clean and whether cell outputs survive. If notebook edits hold up, the rest of the workflow follows.
For R and Julia users, the built-in consoles and environment pane are the draw. These replicate the RStudio experience that generic editors lack. Verify that your packages load and that the help pane resolves your documentation.
What to Try Right Now
Download Erdos from lotas.ai/erdos and open one real notebook. Connect a console. Ask the AI to refactor a single function or fix one plot. Watch whether it reads the live session correctly.
Try it with the bring-your-own-key option first. It avoids vendor lock-in during evaluation and keeps cost predictable. If the notebook editing and console integration hold up on your own data, the tool earns a longer trial. If it does not — you lost an hour, not a workflow.
Information is accurate as of the publication date. Terms, prices, and regulations may change — verify with relevant professionals.