Kilo Desktop bundles 500 AI models, coding agents, and conda

Anaconda has released Kilo Desktop in beta, combining access to more than 500 AI models with coding agents, notebook editing, conda environment management, and local inference. Kilo Desktop is available as a free download for macOS, Windows, and Linux. Hosted model calls remain billable at provider rates.

Anaconda acquired Kilo Code in July 2026. Kilo began as an open-source agentic coding platform, and its agent remains available through VS Code, JetBrains, Cursor, terminal, mobile, and Slack integrations. The desktop application extends that platform into data science workflows and centrally managed development environments.

A model switchboard with cost routing

Many AI vendors tie their agent harness, the software layer that supplies tools and project context, to their own model families. Kilo Desktop provides a common interface for models from frontier labs and open-source providers, including NVIDIA, Moonshot, MiniMax, and Z AI. Developers can also connect self-hosted models and third-party subscriptions through presets or custom endpoints.

Kilo says it charges provider rates without adding a model-usage markup. Its Auto Efficient router classifies each request, consults results from Kilo Bench, the company’s coding benchmark, and selects the lowest-cost model that has demonstrated the required capability. Teams can still choose a specific model when latency, context length, data residency, or reproducibility takes priority.

Four agents, multiple repositories

Kilo Desktop includes four agents with distinct permissions and responsibilities:

• Code writes and edits files. After a plan is approved, it can divide the work into subtasks and run them concurrently through sub-agents.
• Plan examines the project and produces an implementation sequence without modifying files.
• Ask answers questions about the codebase while keeping the project unchanged.
• Debug investigates failures and applies fixes.

Teams can create and share custom agents for internal workflows. Each Kilo workspace can also group multiple folders from the same machine, allowing an agent to work across repositories. A service repository, analysis project, and shared library can therefore sit within one agent context while remaining separate on disk.

Notebook edits meet managed conda

Kilo Desktop can edit notebook cells directly, preserving the notebook structure and existing execution context. Developers and researchers can review generated code where it will run, avoiding the copy-and-paste loop common to chat-based notebook assistance.

The application also creates and manages conda environments. Packages come from Anaconda’s catalog of more than 19,000 reviewed packages, giving teams a controlled source for Python dependencies and isolated runtimes. Integrating environment management with the agent also gives generated code a defined place to execute and makes dependency changes easier to inspect.

Local inference and extensible tools

A built-in model server runs open models on local hardware for code or data that must remain on the machine. Full agent mode gives a local model access to Kilo’s coding tools and project context. A lightweight mode supports simpler inference with lower overhead. Available speed and model size will depend on the machine’s CPU, GPU, and memory.

A marketplace supplies skills, custom agents, and Model Context Protocol servers at either workspace or global scope. MCP is Anthropic’s open protocol for connecting AI agents to external tools and structured data sources. Its inclusion lets Kilo use community-built connectors while giving administrators control over which integrations enter a workspace.

Policy controls inside the client

Administrators can enforce organizational policy through several controls:

• Model allow-lists
• Approved providers and geographic regions
• Agent permissions and available tools
• Permitted skills and integrations
• Local or self-hosted inference options

Embedding these controls in the application gives security teams a consistent enforcement point across models and agents. The approach is relevant to organizations that restrict external model access, require regional processing, or need tighter control over tools that can modify source code.

Anaconda says the broader Kilo platform processes more than 10 trillion tokens each month. Kilo Desktop is a newly released beta, so that usage figure describes the established platform rather than adoption of the desktop client.

A separate ChatGPT integration allows ChatGPT Plus and Pro subscribers to sign in and route eligible plan usage through Kilo, reducing the need to configure another API key.

Best fit for mixed AI stacks

Kilo Desktop has the clearest fit in development environments with one or more of these requirements:

• Frequent switching between frontier and open-source models based on cost, latency, capability, or data location
• Projects that combine application code, notebooks, and shared libraries
• Python workflows that depend on managed conda environments
• Local inference for sensitive code or data
• Central model, provider, region, and agent permission policies

The beta download is free, while hosted inference is charged at provider rates and local inference uses the team’s own hardware. Before production deployment, teams should test repository permissions, notebook fidelity, environment reproducibility, model-routing quality, local performance, and policy enforcement against their existing development and security requirements.