Databench by Alkera

One notebook holds SQL and Python cells side by side. A SQL cell queries any configured connection and hands the result to Python as a dataframe, and tables and charts render inline. Files stay plain Python, compatible with marimo, so they open in an editor and diff in git like any other source file. That matters when a team wants review history rather than a screenshot of a result.

An agent reads, writes, and runs cells in the notebook you have open, and its changes appear as it makes them. Several people can join the same chat, which turns a solo analysis into a shared session. You can stop a run, edit a cell, and send the agent back with a correction instead of restating the whole request.

An organisation admin adds any Linux server reachable over SSH, and Databench installs a node on it. Machines are shared per team, so a GPU box can serve the people who need it without a separate cluster account. Each chat runs in its own sandbox with gVisor and nftables rules that leave other services on the host alone.

Chats reach Anthropic, OpenAI, or Bedrock through a gateway that uses keys you supply. Nothing is resold and no shared quota sits between you and the provider. Swapping a key or a provider is a configuration change on your side, not a support ticket.

Password sign-in covers organisations, nested teams, invitations, and roles. One authorization policy layer classifies every action and records the decision, which gives a reviewer a trail when an agent proposes something destructive. This is the part that makes the open-source data workspace usable with contractors and auditors.

Each organisation gets a shared drive with live editing, sharing, and trash, so a CSV a teammate drops in is available to every notebook in the workspace. The alkera CLI runs a local daemon and talks JSON-RPC to the web app, which lets a developer work from a terminal without losing the shared session.