Cline v4.1.18 and CLI 3.0.62 ship Desktop app and major SDK updates

Cline has released a synchronized suite of updates across its CLI, SDK, and a new native desktop application, marking a significant shift in how the AI coding assistant handles performance and user experience. According to the GitHub release notes, the headline feature is the introduction of Cline Desktop, a native app designed specifically for working with open-weight models. This release coincides with SDK v0.0.83 and CLI v3.0.62, which collectively address long-standing performance bottlenecks and introduce a robust retry mechanism for transient provider errors.

The details

The most substantial technical change in this release cycle is the overhaul of the streaming and execution pipeline. Previously, every streamed token was proxied to client-side hooks as a capability round trip carrying a full session snapshot, causing the agent loop to await each chunk. This resulted in significant latency, with each token costing hundreds of kilobytes of serialization and a blocking IPC hop. In v4.1.18 and CLI 3.0.62, per-chunk text, reasoning, and tool-update deltas are no longer forwarded to remote hooks, drastically reducing overhead. Additionally, the hub event log moved to synchronous NORMAL mode, dropping one fsync per appended delta.

Reliability has been improved through a new retry system. A model turn that dies mid-stream with a transient provider error is now retried up to three times with exponential backoff. Previously, a single forwarded 429 error would abort the run outright. The system reads retryability from the AI SDK's typed signals and ensures that a turn is never retried once it has streamed any text, reasoning, media, or tool call, preventing data duplication. Model calls also now allow five SDK-level retries for request-start 429/5xx/network failures, up from two.

The release also introduces Hub-managed Agent Plugins. Packages under ~/.agents/plugins/* are discovered and validated from their root plugin.json. Valid skills are exposed through the skills tool as plugin-name:skill-name, and MCP servers are started without touching cline_mcp_settings.json. Notably, workspace .agents/plugins directories are deliberately not scanned, ensuring that opening a repository cannot implicitly start repo-controlled MCP servers.

Context

These updates follow a period of rapid expansion in Cline's model catalog. The latest releases refresh the catalog from 5,923 to 6,079 models, adding providers such as Infer by Flow7, Melious, and Wallaby. This wide refresh changes the resolved default model for 33 to 44 providers, with many landing on DeepSeek V4.1 Flash. For instance, Hugging Face, Fireworks, and DigitalOcean now default to DeepSeek V4.1 Flash, while NVIDIA moves to GLM 5.3 Flash. This shift reflects the broader industry trend toward leveraging high-performance open-weight models for cost-effective coding tasks.

The introduction of the Desktop app also signals Cline's move toward a more integrated user experience. The app includes task import from Claude Code and Codex, scheduled runs, web search, and voice input. It also features a marketplace for plugins, MCP servers, and skills, positioning Cline as a comprehensive platform rather than just a CLI tool.

What it means for developers

For developers using Cline, the most immediate benefit is improved performance. The removal of hook forwarding for per-chunk deltas should result in noticeably faster streaming responses. The new retry mechanism means that transient network issues or rate limits from providers like OpenRouter will no longer cause entire runs to fail, reducing the need for manual intervention.

Developers should be aware of the changes to default models. If you are using a provider without pinning a specific model, expect a different default. For example, if you were using Hugging Face and relying on the previous default, you will now be routed to DeepSeek V4.1 Flash. It is recommended to review your configurations and pin specific models if consistency is critical.

The new Agent Plugins system offers a cleaner way to manage MCP servers and skills. By avoiding the scanning of workspace directories, Cline prevents potential security risks associated with repo-controlled MCP servers. Developers can now manage plugins through the Hub, with enablement living in hub settings keyed by manifest name.

What to watch

• Model Default Changes: Monitor your workflows for any unexpected behavior due to the new default models, especially if you are using providers like Hugging Face, Fireworks, or NVIDIA.
• Desktop App Adoption: Observe how the new Desktop app integrates with existing workflows, particularly the task import features from Claude Code and Codex.
• Plugin Ecosystem: Track the growth of the plugin marketplace and the adoption of Hub-managed Agent Plugins, which may become the standard way to extend Cline's capabilities.
• Performance Metrics: Measure the impact of the streaming and checkpointing optimizations on your specific workspaces, especially those with large untracked directories.