Cline v3.35: Native Tool Calling and MiniMax M2

TL;DR

• Native tool calling replaces XML-based tool definitions with JSON schemas, reducing errors and token usage by ~15%
• Auto-approve menu redesigned as an inline expanding interface with cleaner UX and fewer redundant options
• MiniMax M2 available free until November 7 with interleaved thinking for complex multi-step tasks

What Dropped

Cline v3.35 ships three major updates: native tool calling across next-gen models, a completely redesigned auto-approve menu, and free access to MiniMax M2 through November 7. All three are enabled by default.

The Dev Angle

Native tool calling is the headline here. Instead of asking models to output XML-formatted tool calls within text responses, Cline now sends tool definitions as JSON schemas directly to the API. Models return tool calls in their native JSON format—what they were actually trained to produce. This cuts "invalid API response" errors significantly, especially on GPT-5-Codex where the team has seen substantial improvements in operation success rates.

The practical benefits: parallel tool execution (read three files at once instead of sequentially), smaller system prompts (tool definitions moved to API parameters), and approximately 15% token savings per request. Native tool calling works on Claude 4+, Gemini 2.5, Grok 4, Grok Code, and GPT-5 across Cline, Anthropic, Gemini, OpenRouter, xAI, OpenAI-native, and Vercel AI Gateway. Older models fall back to text-based tool calling.

The auto-approve menu redesign removes friction. The new inline expanding interface doesn't obstruct your view, uses gradient backgrounds for visual separation, and eliminates redundant options (if both "Read" and "Read (all)" are enabled, only "Read (all)" shows). The notifications setting moved to General Settings. Cline removed the main toggle, favorites system, "Toggle all" option, and max requests limit—features that added complexity without workflow value.

MiniMax M2 implements "interleaved thinking"—continuous internal reasoning throughout the entire API request, not just upfront. The model re-evaluates its approach as it processes tool outputs and new information. You'll see thinking blocks in the UI. This matters for long refactoring operations or multi-file changes where context drift kills productivity.

Should You Care?

If you're using Cline with GPT-5-Codex or other next-gen models, native tool calling is a direct win: fewer errors, faster execution, lower token costs. The auto-approve redesign is a quality-of-life improvement—less visual clutter, faster approvals.

If you're on older models (GPT-4, Claude 3.5), you'll stay on text-based tool calling. No breaking changes, but you won't see the error reduction or token savings.

MiniMax M2 is worth testing if you're doing complex, stateful coding tasks. The interleaved thinking approach is genuinely different from standard reasoning models. Free until November 7 means zero risk to try it.

Update through your extension marketplace. Full documentation and changelog available.

Source: Cline