AI coding tools
Cline Desktop Moves the Coding Agent Out of the IDE, Using a Shared Runtime for Parallel Work, Scheduling, and Cross-Agent Imports
Cline has released an open-source desktop beta that lets multiple agents run concurrently in isolated sessions and supports unattended tasks through cron scheduling. However, the model results highlighted by the company come from an existing CLI harness, not an independent evaluation of the desktop app.
Cline released an early version of Desktop on September 14, moving its coding agent—which had primarily been available as a VS Code extension and CLI—into a native workspace for macOS and Windows. The desktop app uses the same agent runtime rewritten with the Cline SDK, allowing editor, terminal, and desktop sessions to share tool calling, context management, error recovery, and model interfaces instead of requiring a separate desktop agent implementation.
The new interface primarily addresses the challenge of managing multiple delegated tasks simultaneously. Users can have separate sessions modify code, browse the web, or run routine checks, while scheduled tasks can be configured with a prompt, working directory, model, provider, and recurrence interval. Marketplace centralizes the management of plugins, MCP servers, and skills, enabling agents to connect to databases, APIs, and internal tools. This also means the desktop application is more than an IDE shell: it is a local agent control plane capable of retaining workspace permissions for extended periods.
The model layer is not tied to a single provider. Cline says users can choose from more than 300 models through its service, connect to over 50 providers, or use local endpoints through Ollama, LM Studio, or an OpenAI-compatible API. Different models can be used for the Plan and Act phases. Conversations from Claude Code, Codex, and other supported agents can also be imported and continued with another model. However, what is imported is resumable conversation context; this does not mean the original agent’s execution environment, credentials, or uncommitted state can be reconstructed seamlessly.
The Terminal-Bench 2.0 figures cited in the release post—such as Kimi K3’s 82.02% pass@1—explicitly reuse results from the September 2 CLI harness test and should not be treated as evidence of Desktop’s reliability in parallel or scheduled modes. Engineering teams should pay closer attention to state consistency after interruptions during long-running tasks, tool-format compatibility across models, and whether permissions can be revoked quickly when a scheduled agent goes out of control. In particular, when MCP, shell access, web search, and third-party models are enabled at the same time, the data boundaries for source code, credentials, and tool outputs depend on each provider and plugin. An open-source client does not, by itself, mean that the entire workflow runs locally.