Anaconda Launches Kilo Desktop Beta With Access to 500+ AI Models

Why Anaconda built a single app

Anaconda argues that building with AI usually means juggling several tools. Frontier labs such as OpenAI, Anthropic and Google build models that teams want for their hardest problems, but according to the company, each lab's agent harness is locked to its own model family. That makes it impossible to switch to open-source models from NVIDIA, Moonshot, MiniMax, Z AI and others.

Some teams also want to build on local models, and want more than generic code generation. Notebooks, conda environments and terminals tend to live elsewhere, and a security team has to approve each tool separately. Kilo Desktop is meant to put that work in one place, inside a tool security teams can approve.

Workspaces, agents and notebooks

Anaconda says real projects rarely sit in one folder, for example an API and the analysis behind it in separate repositories. A Kilo Desktop workspace is a named group of folders from the user's machine, so the agent can read the analysis while it writes the service, without copying anything between them.

The app ships with four agents:

• Code writes and edits files.
• Plan reads the project and returns a step-by-step implementation plan without changing anything.
• Ask answers questions about the code without changing it.
• Debug finds what broke and fixes it.

Once a plan is approved, Code takes over, splits it into subtasks and runs them in parallel with sub-agents. Users can also build custom agents and share them with their team.

For data work, the agent edits notebooks directly rather than through pasted cells. Users can watch its cursor move through the cell as it writes and review the work in the notebook with the data loaded. Kilo Desktop also creates and manages conda environments inside the app, drawing on Anaconda's catalog of more than 19,000 vetted packages.

The interface treats chats, notebooks, files, diffs and terminals as tabs that can be arranged in groups. A carousel mode handles wide layouts, and any tab can go full screen or pop out onto a second monitor.

Model choice, local models and cost routing

The model picker covers every major frontier lab, leading open-source models and providers from every geography, according to Anaconda. Users who do not want to choose can rely on Kilo's auto routers. Auto Efficient checks a request against Kilo Bench, the company's own coding benchmark, and sends it to the cheapest model that has proven capable of that kind of task.

Models users host or subscribe to elsewhere can be added through a preset or a custom endpoint with their own API key. A built-in local model server runs open models on the user's own hardware, so prompts stay on the machine. One setting decides whether local requests carry Kilo's full agent toolkit, which suits work on code, or go without it, which makes them faster and lighter.

The Kilo marketplace offers skills, Model Context Protocol (MCP) servers and custom agents. These can be installed for one workspace or for all of them.

Governance for enterprises

Anaconda says breadth without governance is a liability. Administrators can set which models people may use, which providers and geographies are allowed, what each agent may do and which skills it can install. Because the controls live in the product, the policy a security team writes is the one the app enforces, the company says.

Kilo Desktop is available to download now as a beta. For teams that want several frontier and open-source models, local options and Python environments under one managed roof, it is Anaconda's attempt to remove the need to stitch separate tools together.