Working VS Code AI agent in your sidebar – that’s where this ends. Cline v3.81, reading your repo, editing multiple files at once, running terminal commands, waiting for your approval at every step. Here’s how to get there.
Cline is the agent. Per the official repo, it’s an autonomous coding agent available as an SDK, IDE extension, or CLI assistant. It’s Apache 2.0, has 5M+ installs and 61.2k GitHub stars, and currently ships at v3.81 (as of late-2026 release notes – check the GitHub releases page for newer builds). The VS Code extension ID is still saoudrizwan.claude-dev because the project originally launched as “Claude Dev”.
System requirements
Minimal. One RAM note worth reading before you start, though – skip ahead to the OOM error section if you’re on an older build.
- VS Code: A recent build (the extension manifest pins specific VS Code APIs; check the marketplace listing for the current minimum)
- OS: macOS, Linux, or Windows 10/11
- RAM: v3.80 raised the Node.js heap to 8 GB to stop OOM crashes – see the OOM section below for what that means on lower-RAM machines
- Node.js: Only needed if you also want the CLI (
npm i -g cline) - An API key from a supported provider, or a local model server
No GPU, no Docker, no container runtime required for the extension itself. Local models via Ollama need their own RAM budget on top of whatever VS Code is already using.
Install the VS Code AI agent in three steps
Skip the marketplace search box – install by ID to avoid typo-squatters. In a terminal:
code --install-extension saoudrizwan.claude-dev
Or, inside VS Code:
- Open the Extensions panel (Ctrl/Cmd + Shift + X)
- Search
saoudrizwan.claude-dev– select Cline by saoudrizwan and click Install - A robot icon appears in the Activity Bar. Click it to open the side panel.
That’s the install. Configuration is where most people get stuck.
First-time configuration
Click the gear icon in the Cline sidebar. Pick a provider, paste a key – that’s the minimum viable config.
Per the official site, Cline supports 30+ LLM providers – you bring your own key, endpoint, or weights. The four most common setups (model names as of mid-2026; check each provider’s current offerings):
| Provider | Where to get a key | Notes |
|---|---|---|
| Anthropic | console.anthropic.com | Claude Sonnet 4 / Opus – strong for multi-file agentic work |
| OpenAI | platform.openai.com | GPT-4o and current API models |
| OpenRouter | openrouter.ai | One key, all major models, easier billing |
| Ollama (local) | n/a – install Ollama | Serves at http://localhost:11434; free but slower |
Paste the key, save, then confirm two things: the workspace root is correct, and approval-before-execute is on. If you skip the workspace root step, the agent can edit files outside your project – including dotfiles you didn’t intend to touch.
Optional but worth doing: drop a .clinerules file at your repo root with project conventions. The official README notes these rules – coding standards, architecture conventions, deployment procedures – get picked up automatically by the CLI, VS Code extension, and JetBrains plugin alike.
Verify the install works
Open any project folder. Click the Cline icon. In Plan mode, type:
What does this codebase do? List the top-level files and their roles.
If it reads files and produces a summary within ~30 seconds, you’re live. Toggle between Plan mode and Act mode – in Plan mode, Cline explores your codebase, asks clarifying questions, and lays out a strategy; once you’re aligned, switch to Act mode and it executes the plan, with every file edit and terminal command requiring your approval.
Pro tip: Start every task in Plan mode, even simple ones. The plan is free (no tool calls), and reading it before flipping to Act catches bad assumptions before they become destructive edits. Reviewing a wrong plan costs nothing; reverting wrong files costs your afternoon.
That approval loop is the thing that takes getting used to. You’re not watching an AI work – you’re actively directing it. Some people find this tedious at first. After a few sessions it starts feeling less like supervision and more like having a junior dev who actually asks before doing something irreversible.
Common install errors and the fixes that actually work
Day one usually breaks in one of these ways. Pick your symptom.
1. Agent dies mid-task on long sessions (OOM)
Symptom: extension host crashes or sidebar goes blank after a long Plan/Act cycle. The root cause: before v3.80, Node.js heap limits were too low for long sessions. The changelog confirms v3.80 raised the heap to 8 GB specifically to fix this – so if you’re on an older version, that’s your problem. By default VS Code age-gates new Cline releases on auto-update. Skip the wait:
# Force-fetch latest published version from the marketplace
code --install-extension saoudrizwan.claude-dev --force
Manual installs grab the latest published version immediately, bypassing the age gate that delays automatic updates.
2. Terminal output never reaches the agent
Symptom: Cline runs npm test, you see the output in the terminal, but Cline insists “no output received.” Shell integration failure – usually a custom prompt or non-standard shell. The changelog records a fix that landed as a fallback capturing raw terminal content directly when shell integration fails, so updating to current main resolves most cases. If it persists, switch your VS Code default profile to plain bash/zsh and reload.
3. ENAMETOOLONG when using Claude Code provider
Long conversation histories blow past the OS argv limit on Linux/macOS. This has been patched – but if you’re on a stale build, start a new task instead of continuing an old one. Each new task starts a fresh argv.
4. “Provider not found: gemini-cli” after upgrade
Had Gemini CLI configured previously? It’s gone. The changelog notes Google asked them to remove it. Switch your Gemini config to the standard Gemini API key flow instead.
5. Enterprise: Azure Blob Storage backend auth failure
Almost never documented anywhere. For v3.79+ enterprise users, if you’re using Azure Blob Storage as the storage backend, confirm those credentials separately. The extension assumes ambient Azure auth that may not be present in your environment.
Upgrade and uninstall
Upgrades are automatic through VS Code’s extension auto-update, but age-gated as noted above. To force the newest published build right now:
code --install-extension saoudrizwan.claude-dev --force
Clean uninstall – extension plus stored state:
- Extensions panel → Cline → cog icon → Uninstall
- Delete the extension’s storage folder:
- macOS:
~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev - Linux:
~/.config/Code/User/globalStorage/saoudrizwan.claude-dev - Windows:
%APPDATA%CodeUserglobalStoragesaoudrizwan.claude-dev
- macOS:
- If you installed the CLI:
npm uninstall -g cline - Revoke any API keys at your provider console – uninstalling locally doesn’t invalidate them
That last step matters more than people think. The keys keep working from any machine until you rotate them.
What to do next
Open a real project. Switch to Plan mode. Ask Cline to add a single endpoint, a single test, or refactor one function – something small enough that you can read every diff before approving it. The trust calibration takes about three tasks. After that you’ll know which jobs to delegate and which to keep manual.
FAQ
Is the Cline extension free?
Yes – free and open-source (Apache 2.0). You bring your own API key for Claude, GPT, Gemini, or any other supported LLM, and pay only for the API usage directly. No Cline subscription.
Can I use Cline fully offline with a local model?
Yes – point it at Ollama or LM Studio in the provider dropdown. Concretely: install Ollama, run ollama pull llama3, set Cline’s base URL to http://localhost:11434. The tradeoff is real: smaller local models handle targeted edits fine but can struggle with multi-file refactors that require holding a lot of context. A common pattern is using local models for routine edits and cloud models for harder tasks – but your hardware and privacy requirements will dictate what makes sense for you.
What’s the difference between Cline and the Cline CLI?
The VS Code extension runs in your editor with a sidebar UI. The CLI (npm i -g cline) runs the same agent core headlessly – useful for CI pipelines or scripts where you don’t want a GUI. They share .clinerules and most config, so a repo configured for one works with the other.