Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/Kilo desktop for data science and research is now generally available
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

Kilo Desktop for Data Science and Research Is Now Generally Available

Фото: Kost9n4 (Pixabay) — https://pixabay.com/photos/analysis-biochemistry-biologist-2030265/

Source: Anaconda•November 18, 2026 at 08:55 PM

Kilo Desktop for Data Science and Research Is Now Generally Available

Kilo Desktop for data science is now generally available. Run an AI agent in live Jupyter notebooks, manage conda environments, and use governed packages.

As a data scientist or researcher, you spend much of your working day in a notebook. When an AI assistant can’t access that notebook, you end up copying cells into a chat and answers back into your analysis. Kilo Desktop is now generally available, bringing an agent into live Jupyter notebooks to edit and run cells, display data files, and manage environments and packages, including packages from Anaconda’s governed channels. If you already use Anaconda, import an environment.yml and start from the setup you already have. More than 5 million developers already build with Kilo.

Kilo Desktop also creates interactive charts, dashboards, and data tables you can save to your workspace from the conversation, with more than 500 models to choose from, including ones on your own hardware. If you tried an earlier beta, recent updates bring multi-agent sessions, the Workspace Board, and the Anaconda Local Model Catalog.

Let’s dig deeper: Notebooks, environments, and packages in one place

Open any .ipynb file, or ask the agent to create a new one, and it opens as a live notebook beside your chat. You and the agent share the same document: you watch it write a cell, run it, read the output, and move on. The agent has access to the same results you do, plots included, so when a chart looks wrong, it can flag the problem.

Because it works in a real kernel, the agent handles the unglamorous parts too: reading tracebacks, diagnosing a kernel that won’t start, restarting a stuck one. Ask it to profile a dataset, try three feature-engineering ideas, or turn a messy exploration into a clean notebook, then review the result with the data still loaded.

You approve code before it runs

Running code changes your kernel’s state, so Kilo Desktop prompts you first. In Manual mode, running a cell and restarting the kernel each show an approval card. “Always allow” covers that notebook for the rest of the chat. “Automatically approve” approves these prompts on your behalf, so enabling it is a trust decision.

Adding packages has its own card. It lists the new packages plus any packages the notebook already declares that aren’t approved yet, and flags names that look like misspellings of popular packages (numppy for numpy, requets for requests). Package approvals are machine-wide: once approved, a package can be used in any notebook on the computer. Pre-approved packages such as pandas, numpy, and matplotlib don’t trigger a prompt, and flagged names can’t be remembered with “Always allow.”

Conda environments without leaving the app

You and the agent can both create and manage conda environments inside the app. Start fresh with a Python version, import an environment.yml, or clone an environment before trying a new approach with a different package. You review dependency changes before they happen, and Back up and update clones the environment first, so a bad upgrade can be undone. New notebooks get a ready environment with ipykernel automatically, so the first cell runs without needing additional installs.

Anaconda’s governed packages, by default

When you sign in, the agent installs from Anaconda’s catalog of more than 19,000 vetted Python packages and any governed channels your organization has set up. Its install tools reject packages from local files and URL channels, and its shell commands are permission-gated. Your Anaconda sign-in shares the keyring of the conda command-line interface (CLI), so authenticated channels work inside and outside the app.

From first look at the data to a finished chart fast

CSV, TSV, and Parquet files open as tables in the Files pane, so you can spot problems before you ask a question. @-mention a file so the agent works from the file you mean, or open it in an editor you’ve installed on your PATH.

Interactive charts in the chat

Some answers are easier to read as a chart. The agent draws interactive Plotly, Bokeh, and HoloViz charts and data tables right in the conversation, styled to match your theme. Save one to your workspace or open it full size in the viewer. Diagrams render inline too.

A layout built for analysis

Chats, notebooks, files, diffs, terminals, and web pages all open as tabs you can arrange into groups. Keep the notebook beside the chat, serve a Panel or Streamlit dashboard from the terminal and preview it in the built-in browser, and keep the Changes pane open to review the files the agent changed. Panes can pop out onto a second monitor. Save your favorite arrangement as a layout preset (for example, chat, notebook, and data side by side), and new sessions open that way.

A workspace can span several folders, so the agent reads your analysis, pipeline code, and shared utilities together. Run speculative experiments in isolated git worktrees, and when the work is ready, Kilo writes the commit message and pull request.

Plan first, then build

Plan reads your data and code and returns a step-by-step approach without changing anything. Approve it, and Code carries it out, splitting the work across sub-agents. Ask explains an unfamiliar notebook, and Debug tracks down why last night’s numbers changed.

While a long cell runs, queue your follow-ups (for example, “now break it down by region”), then reorder or edit them before they run. Kilo Desktop notifies you when the agent finishes or needs your input. Juggling several analyses? The Workspace Board lays your sessions out in columns you name, like Exploring, Modeling, and Reporting, so you can see where each question stands at a glance.

Models you choose, including ones on your own hardware

Models from frontier labs such as OpenAI and Anthropic, and open models such as NVIDIA Nemotron, are available from one picker. If you’d rather not choose, Auto Efficient checks your request against Kilo Bench and sends it to the least expensive model proven capable of that kind of task. Already have an API key, an Amazon Bedrock account, or a self-hosted endpoint? Add one as a provider and use it from the same picker.

Local models for sensitive data

Some data can’t leave your machine. The built-in Local Model Server runs inference for open models on your hardware, without sending those model requests to a hosted model provider. Browse the Anaconda Local Model Catalog for models that fit your hardware, or import GGUF models you already downloaded with Anaconda Desktop.

Skills, MCP servers, and custom agents

The Kilo Marketplace carries skills, Model Context Protocol (MCP) servers, and custom agents you can install for one workspace or all of them. Plug in your data warehouse or your team’s notebook conventions once, and new projects start with them.

Why this matters: Built for practitioners, ready for your organization

On Kilo Enterprise, organization owners can block specific models and providers. Combined with Anaconda’s governed channels, your data team gets an agent that works within the rules your organization has already set. Kilo Desktop is part of Anaconda’s AI Workspaces, with packages and models from AI Artifacts.

Give it a shot

If you’ve been wanting to use AI tools in your analysis and research, give Kilo Desktop a try. Download Kilo Desktop, open a notebook, and ask it to help explore a dataset. Review the code, run the analysis, and save your first chart.

← All articles

What this article says

More in Software Development

All →
Cyberattacks are getting faster. Here’s how Microsoft is fighting back
Microsoft

Cyberattacks are getting faster. Here’s how Microsoft is fighting back

Microsoft unveils 7 AI models for coding, image editing, voice and more
Microsoft

Microsoft unveils 7 AI models for coding, image editing, voice and more

Inside Secretless Analysis: from CI identity to project-scoped access
SonarSource

Inside Secretless Analysis: from CI identity to project-scoped access

Java4Shell (CVE-2026-114938): Find affected Java in your container images
SonarSource

Java4Shell (CVE-2026-114938): Find affected Java in your container images

10 Questions With August Tang on Interfaces and Internet Nostalgia
Squarespace Commerce

10 Questions With August Tang on Interfaces and Internet Nostalgia

Is Apple the safe haven investors think it is? Here's a way to trade it with less riskПресса
Apple

Is Apple the safe haven investors think it is? Here's a way to trade it with less risk

More from Anaconda

Help Your Team Code Faster and Explore Further with Kilo Desktop
Anaconda

Help Your Team Code Faster and Explore Further with Kilo Desktop

AI Red Teaming: What to Test and How to Act on Findings
Anaconda

AI Red Teaming: What to Test and How to Act on Findings

A Fintech Giant Built Across the Americas. Can Your AI Team?
Anaconda

A Fintech Giant Built Across the Americas. Can Your AI Team?

Only 22% of Security Teams in Our Live Tabletop Had a Tested AI Agent Runbook
Anaconda

Only 22% of Security Teams in Our Live Tabletop Had a Tested AI Agent Runbook

Enkrypt AI by Anaconda Named a Sample Vendor for AI Security Testing in Two Gartner® Hype Cycle™ Reports
Anaconda

Enkrypt AI by Anaconda Named a Sample Vendor for AI Security Testing in Two Gartner® Hype Cycle™ Reports

Anaconda Joins the Open Secure AI Alliance to Advance Secure AI Development
Anaconda

Anaconda Joins the Open Secure AI Alliance to Advance Secure AI Development