v2.1.0

Cline now uses Anthropic's new "Computer Use" feature to launch a browser, click, type, and scroll. This gives him more autonomy in runtime debugging, end-to-end testing, and even general web use. Try asking "look up the weather in Colorado" to see it in action! (Available with Claude 3.5 Sonnet v2)

See a demo here! https://x.com/sdrzn/status/1850880547825823989

v2.0.0

• New name: Meet Cline, an AI assistant that can use your CLI aNd Editor. While “Claude Dev” was a tribute to Claude 3.5 Sonnet, v2.0 brings updates that significantly improve performance across other models making it a good opportunity to give the extension its own identity.

• No more staring at loading spinners: Responses are now streamed into the editor + a yellow text decoration animation to keep track of Cline’s progress as he edits files. Watching Cline work feels like magic and I’m so incredibly excited for you all to experience it.
  

• Cancel button: With the new streaming feature, you can now Cancel mid-generation to give Cline feedback if he goes off in the wrong direction. This gives you more control over tasks, leading to the results you want faster.

• 40% fewer requests to accomplish tasks + better performance with other models: This update replaces the rigid tool use API with a customized XML tags approach, allowing for more flexible tool interactions and significantly improving results for non-Claude models. Behind the scenes, older versions of the extension used an API feature called “Tool Use” (AKA “Function Calling”) where tools like read_file and execute_command were defined in a strict JSON schema sent with each request, requiring models to respond with a structured output that had to fit into this schema. While some models like Claude 3.5 Sonnet are heavily fine-tuned to be good at this, other models like Llama and DeepSeek are not as adept, often leading to inability to complete even the most basic tasks. This not only limited the extension’s accessibility with other models, but also diminished Claude 3.5 Sonnet’s potential as forcing any model to respond in structured output has been proven to have negative qualitative effects. To eliminate these bottlenecks while still giving models the ability to invoke tools, I took inspiration from this study showing how prompt-engineering can enable stable tool-calling capabilities + Anthropic’s prompt to achieve their Artifacts feature (uses tags like <antArtifact>) to create a completely re-imagined system prompt where models are effectively fine-tuned on the spot how to best interact with Cline’s toolset. Instead of having to respond in a rigid JSON format, models can now invoke tools in unstructured text with XML tags like <read_file> and <execute_command>. Since we define the tool use guidelines ourselves we can even tailor it with context like “these tools are executed upon the user's approval” instead of relying on the generalized internal prompts that Anthropic uses to achieve tool calling for their API. To quantify the improvement these changes made, I ran a simple benchmark: I gave the following task in a Next.js app “Implement a dark mode toggle feature, and save the user’s preference to persist across sessions” and measured the # of requests until the task was successfully completed. This was a practical scenario that tested models’ abilities to use the appropriate tools to explore an existing project’s source code, apply edits, and debug issues that came up along the way.
  
  Claude 3.5 Sonnet saw a ~40% reduction in # of requests required to accomplish the task, while other models like Llama and DeepSeek are now not only able to tool call more reliably but seem to have a greater capacity to reason about executing tasks successfully.
  While the native tool calling API was a core feature of earlier versions of the extension, it’s become clear that it limited the models' performance. The more flexibility the models have in generating responses, the better the outcomes. Although it’s more work on the extension’s part to parse the XML tool invocations, it is certainly worth the trouble.

• Search and use any OpenRouter model: the extension automatically fetches the latest model list + pricing details, giving you access to the newest models as soon as they’re available. Try searching “free” to find no-cost options.

• Apache 2.0 License: Changed the project’s license from MIT to Apache 2.0 to provide stronger open-source protections for the future development of the project. Thank you contributors for making this a smooth process.

• New README and got a cool domain for the project https://cline.bot (wip). On a final note, I also wanted to share that I’m now working on Cline full-time and am excited to keep improving the extension and make it the best thing it can possibly be. Thank you all for the incredible support around the project.

–Saoud

v1.9.0

• Claude can now use a browser! This update adds a new inspect_site tool that captures screenshots and console logs from websites (including localhost), making it easier for Claude to troubleshoot runtime/visual issues on his own. See a demo here.
  
• Improved automatic linter/compiler debugging by only sending Claude new errors that result from his edits, rather than reporting all workspace problems.

v1.8.0

You can now use @ in the textarea to add context!

• @url: Paste in a URL for the extension to fetch and convert to markdown, useful when you want to give Claude the latest docs!
• @problems: Add workspace errors and warnings for Claude to fix, no more back-and-forth about debugging
• @file: Adds a file's contents so you don't have to waste API requests approving read file (+ type to search files)
• @folder: Adds folder's files all at once to speed up your workflow even more

v1.7.0

Claude Dev can now monitor your workspace for linter, compiler, and build issues as he works 🚀 This lets him automatically fix problems like missing imports, type errors, and more all on his own!

Check out a demo here!

Update: This feature is now opt-in in v1.8.0 (use @problems in the textarea). Automatic debugging was an ambitious demand for Claude 3.5 Sonnet, but I hope to add it back in once models are good enough at fixing hard problems! See this comment for more details.

Update 2: I've added auto-debugging back into v1.9.0 but now the extension only automatically sends new errors that result from Claude's edits, rather than reporting all workspace problems. This gives Claude autonomy over fixing errors himself while not distracting him with issues unrelated to the task.

v1.6.5

• OpenRouter now supports prompt caching! (They're still working on reporting cache usage, but the API cost will be accurate)
• Adds support for Azure OpenAI and Google Gemini (free for up to 15 requests per minute!)
• Task header can now be collapsed to provide more space for viewing conversations
  
• Adds fuzzy search and sorting to Task History, making it easier to find specific tasks
  

v1.6.0

Commands now run directly in your terminal thanks to VSCode 1.93's new shell integration updates! Plus a new 'Proceed While Running' button to let Claude continue working while commands run, sending him new output along the way (i.e. letting him react to server errors as he edits files).

v1.5.27

• Claude's changes now appear in your file's Timeline, allowing you to easily view a diff of each edit. This is especially helpful if you want to revert to a previous version. No need for git—everything is tracked by VSCode's local history!
  
• Updated system prompt to keep Claude from re-reading files unnecessarily

v1.5.19

• Adds support for OpenAI compatible APIs (e.g. Ollama!)

v1.5.13

• New terminal emulator! When Claude runs commands, you can now type directly in the in-line terminal (+ support for Python environments)
• Adds search to Task History