Cline v4.0.12 shipped on July 28, 2026, and it changed the entry barrier for the most popular open-source coding agent. The headline: free built-in models that work without API keys. Combined with Claude Opus 5 support across every major provider (added in v4.0.11) and a standalone desktop app (v0.0.7), Cline became the easiest coding agent to try — and one of the most capable.
The follow-up releases — v4.1.4, v4.1.5, and v4.1.6, all shipped August 5–6 — add something beginners should know about before anything else: plan mode can no longer touch your files. This page covers both the free-access story from v4.0.12 and the new safety changes that make Cline a safer first coding agent.
If you have been waiting to try an autonomous coding agent but did not want to set up API keys or pay money, this release line is for you.
Plan Mode Now Blocks File Edits — The Big v4.1 Change
If you are new to Cline, the first thing to understand is Plan/Act mode. In Plan mode the agent can look around — read files, search your codebase, run read-only commands — and propose a plan, but it should not change anything yet. In Act mode it actually edits files, runs commands, and gets the job done. The idea is a checkpoint: you review the plan before the agent touches anything.
Before v4.1.4, that checkpoint had two holes. First, the model could decide mid-conversation to switch itself from Plan to Act. Second, plan mode tried to stop file-editing commands with prompts and warnings — which works most of the time but is not a guarantee.
Cline v4.1.4 closes both holes:
- File-editing shell commands are hard-blocked in plan mode. File manipulation, in-place editors (
sed -i,perl -i), redirection into files (> file), mutating git subcommands (git reset --hard), and package installs are refused outright. Read-only investigation still works — the agent can stillcat,grep, and plan. - Model-initiated plan-to-act switching is removed. Only you can switch modes now. The model cannot quietly promote itself from “planning” to “editing.”
For a beginner, this is the most important safety change Cline has shipped in months. When you are still learning what an autonomous agent will do, you want the agent to be physically unable to start editing before you say go. That is exactly what plan mode is now.
The release notes also fix a related annoyance: a turn that ends with a blocked command no longer counts as a failed turn. You see the plan, review it, and switch to Act yourself — and Act re-runs the plan instead of appearing to do nothing.
v4.1.5 and v4.1.6: Clearer Limits, Consistent Reasoning
The two follow-up releases refine the experience rather than add new features, but two changes matter for beginners:
- Free-model notices explain themselves. When a free model promotion ends, you now get a dedicated notice with a button to pick another model, instead of a confusing generic error. The same clarity applies when you hit a free-tier limit — Cline tells you when it resets.
- Reasoning settings behave the same on every provider. Effort levels and “reasoning on/off” toggles now map to a shared path across all AI SDK providers — including local setups like Ollama — instead of each provider interpreting them differently. If “high effort” works on one model, it works on the next one you switch to.
v4.1.4 also hardened reliability in ways you will only notice when things go wrong: running out of context now compacts the conversation and retries once instead of failing with a raw provider error, and empty model responses are retried on every provider, not just Ollama. Both were common “the agent just died” experiences for newcomers.
Free Cline Models: Zero API Keys Required
The biggest change in v4.0.12 is free models. Cline now shows models labeled “(free)” directly in the model picker. No API key setup. No credit card. No subscription tier. You pick a free model from the dropdown and start coding.
When you hit the free-tier usage limit, Cline displays a dedicated error card with the reset time — so you know exactly when you can use it again instead of guessing or troubleshooting.
For beginners, this is a game-changer. Most coding agents require you to already have an API key from Anthropic, OpenAI, or another provider before you can do anything. Cline’s free models remove that friction entirely. Install the extension, pick a free model, describe your task, and the agent starts working.
The free models are a subset of what Cline supports through its provider ecosystem. They are not meant to replace paid frontier models for heavy-duty work — for complex multi-file refactors or architectural reasoning, Claude Opus 5 or GPT-5 will still outperform. But for learning the tool, prototyping ideas, small refactors, and understanding how autonomous agents work, the free tier is more than enough.
Claude Opus 5 Across Six Providers
Cline v4.0.11 added Claude Opus 5 support — Anthropic’s newest and most capable model for coding tasks. What makes this significant is not just the model itself but the breadth of providers that now support it through Cline.
Opus 5 is available through:
- Anthropic (direct API)
- Claude Code (Anthropic’s own agent integration)
- AWS Bedrock
- Google Cloud Vertex
- Cline Provider (app.cline.bot, pay-as-you-go)
- OpenRouter
Each provider includes 1M context window variants. The 1M context means Cline can work with massive codebases — read your entire project structure, multiple large files, and conversation history without losing track of what it is doing.
Why multiple providers matters for beginners: you are not locked into one billing relationship. If your company uses AWS, run Opus 5 through Bedrock. If you already have an OpenAI account, route through OpenRouter. If you want the simplest setup, use the Cline Provider. Same model, same agent, different billing.
Opus 5 costs $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8. The safety classifiers intervene roughly 85% fewer times than Fable 5, meaning fewer interrupted sessions and smoother autonomous runs. For coding agents specifically, this matters because long autonomous sessions need tools and file access without constant permission prompts.
v4.0.11 also added Moonshot Kimi K3 support, with native tool calling enabled. Kimi K3 is a strong alternative for users who want a non-US model option with competitive coding performance.
Standalone Desktop App
Cline Desktop v0.0.7 shipped on July 29. This is the standalone version of Cline that runs outside VS Code — no editor extension required.
The desktop app includes:
- System tray icon showing app status and active session count
- Session history with pagination, favorites, and activity-based ordering
- Subagent and teammate run visibility — see background agents and their results
- Free Cline models — the same “(free)” models from the VS Code extension work here too
- Polished chat UI — tool-specific icons, elapsed thinking time display, anchored message actions
The desktop app also fixed regional routing for Qwen, Moonshot, Z AI, and MiniMax — the China/international endpoint toggles now work correctly, so users in Asia get routed to the right servers instead of silently hitting the wrong host.
Why this matters: the VS Code extension is powerful but tied to one editor. The desktop app works alongside any editor — VS Code, JetBrains, Neovim, Zed, or none at all. It is also the path to using Cline’s CLI for headless automation, CI pipelines, and scripts.
Desktop v0.0.10 (August 7) adds OAuth for remote MCP servers, which is beginner-relevant because MCP is how Cline talks to external tools and data sources. You can now authorize a remote MCP server right from Settings → MCP, see its auth status, and cancel or retry a pending authorization. Servers that need a pre-registered client ID/secret are supported too, and stored tokens are invalidated when a server’s client configuration changes. MCP errors also now appear on the individual server card instead of as a page-level error — so when something is misconfigured, you can see which server is at fault instead of hunting for a vague failure.
CLI Reliability Fixes
v4.0.12 fixed a bug where Claude Code responses that were already streamed would be discarded when the CLI exited with a max-turns error. Before the fix, if the agent hit the turn limit mid-response, you lost the work it had already done. Now the response is preserved — you keep the code changes even if the session ends abruptly.
This is the kind of fix that matters most when you are not watching the agent run. Headless sessions, cron jobs, and CI pipelines rely on the agent finishing gracefully or failing informatively. Losing a streamed response is a silent data loss that can cost hours of work.
How to Get Started
VS Code extension:
code --install-extension saoudrizwan.claude-devThe extension ID is still saoudrizwan.claude-dev from the original “Claude Dev” name.
CLI:
npm install -g cline
cline auth
clineDesktop app: Download from the Cline Desktop releases — macOS only for now.
After installing, open the Cline panel, click the gear icon, and pick your model. For free, choose any model labeled “(free)”. For paid, enter your API key from your preferred provider.
What This Means for the Coding Agent Landscape
Cline is not the only agent offering free or cheap access. Opencode, Pi, and Hermes all have free tiers or open-source options. But Cline’s combination of free models, 8M+ installs, 61K GitHub stars, and a mature feature set (Plan/Act mode, browser automation, MCP, checkpoints) makes it the lowest-friction entry point for someone who has never used a coding agent.
The Opus 5 support also narrows the gap with Claude Code. Cline can now run the same model that Claude Code uses, across the same providers, with the same context window. The difference is in workflow: Claude Code is terminal-native and excels at headless automation. Cline is IDE-native and excels at visual review with diff previews and checkpoint restores. Both are Apache 2.0 or equivalent. Neither locks you in.
For teams evaluating coding agents, Cline v4.1 makes the “try it on a real project” test free — and safer. No procurement cycle, no API key budget request, no credit card. Pick the free model, keep the agent in plan mode while you get comfortable, and see what it does with your codebase.
Related Reading
- What Developers Say About Claude Code vs Cursor
- Claude Opus 5 Release: What It Means for Claude Code
- Best Coding Agents 2026 Decision Guide