Install Continue v2.0.0: Open Source Copilot Setup Guide

Here’s the awkward truth about the most popular open source Copilot: Continue was acquired by Cursor in mid-June 2026, and the continuedev/continue repository is now read-only after a final 2.0.0 release of the VS Code extension, CLI, and JetBrains plugin. You’re installing a snapshot. No new fixes are coming. But the code is Apache 2.0, the extension still ships from the marketplace, and paired with Ollama it still gives you private, local, no-subscription code assistance.

This guide gets Continue v2.0.0 running with a working local model in about 15 minutes. It also covers the version-lag bug in the VS Code Marketplace that will silently give you a stale build, and the sqlite3 error you’ll hit if you try to build from source.

System requirements

Continue is a thin coordinator – the heavy lifting is done by whatever LLM you plug in. So requirements split into two: the extension host and the model backend. The numbers below are estimates based on typical Ollama usage patterns; your exact needs will vary by model choice.

If you skip Ollama and use a cloud API key (Anthropic, OpenAI, Mistral), the local RAM/VRAM row goes away and you’re back to a normal editor install.

The VS Code install (and the version trap)

The obvious path: open Extensions, search “Continue”, click Install. That works. What isn’t obvious is that the VS Code Marketplace has been serving what looks like a positively ancient version – users report the extension claiming to be 1.2.14 while GitHub releases went from 1.3.22 to 1.3.30, and the marketplace showed 1.2.14 as updated “2 weeks ago”. This is a real reported bug documented in GitHub discussion #10178.

To make sure you get the actual final v2.0.0 build:

1. Open the Continue page on the Visual Studio Marketplace in your browser first – check the version string there.
2. If it shows anything below 2.0.0, don’t install via the marketplace. Instead grab the VSIX from the GitHub releases page.
3. In VS Code: Extensions view → ... menu → Install from VSIX… → pick the file.

After install, the Continue logo appears in the left sidebar. Drag it to the right sidebar – the chat panel is more useful there because it doesn’t fight the file explorer for space.

The CLI install (recommended for JetBrains users)

The Continue team recommends the CLI over the JetBrains plugin (per the official docs as of mid-2026), so JetBrains users especially want this. Three install methods:

# macOS / Linux - shell installer (bundles its own runtime)
curl -fsSL https://raw.githubusercontent.com/continuedev/continue/main/extensions/cli/scripts/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/continuedev/continue/main/extensions/cli/scripts/install.ps1 | iex

# Cross-platform if you already have Node.js 20+
npm i -g @continuedev/cli

# Verify
cn --version

The npm package @continuedev/cli was at 1.5.47 as of early July 2026 – this may have moved since, but not by much given the repo status. The binary is called cn.

First-time configuration with Ollama

The reason to run Continue at all instead of GitHub Copilot is usually privacy. Here’s the minimum config that gets you fully local autocomplete + chat.

First, install Ollama and pull two models. Autocomplete and chat have different jobs:

# Install Ollama (macOS)
brew install ollama
# Or Linux
curl -fsSL https://ollama.ai/install.sh | sh

# Autocomplete - must be a FIM-capable coder model, kept small for latency
ollama pull qwen2.5-coder:1.5b

# Chat - bigger, slower, smarter
ollama pull qwen2.5-coder:7b # ~8 GB VRAM

# Sanity check
curl http://localhost:11434 # → "Ollama is running"

Now open the config. In VS Code, Ctrl+Shift+P → Continue: Open Config. This creates and opens ~/.continue/config.yaml (or %USERPROFILE%.continueconfig.yaml on Windows). Paste:

name: Local Copilot
version: 0.0.1
schema: v1
models:
 - name: Qwen Coder 7B Chat
 provider: ollama
 model: qwen2.5-coder:7b
 apiBase: http://localhost:11434
 roles:
 - chat
 - edit
 - apply
 - name: Qwen Coder 1.5B Autocomplete
 provider: ollama
 model: qwen2.5-coder:1.5b
 roles:
 - autocomplete
 defaultCompletionOptions:
 keepAlive: 1800

The keepAlive: 1800 matches the Ollama default of 1800 seconds – how long the model stays loaded after the last request. Drop it lower and cold-start latency kicks in every time you pause coding for more than a few minutes.

Do not put GPT-5 or Claude in the autocomplete role. Per the Continue autocomplete docs: autocomplete models use a specific FIM prompt format (fim_prefix / fim_suffix / fim_middle tokens) that GPT-5 and Claude were not trained with. They won’t generate useful completions. Use qwen2.5-coder, starcoder2, or deepseek-coder instead – most solid autocomplete models are 10B parameters or smaller anyway, which is why latency stays manageable.

Verify it works

1. Chat: open the Continue sidebar, type “hello”. A response streams back. If it hangs, Ollama isn’t reachable – recheck curl http://localhost:11434.
2. Edit mode: highlight a function, press Cmd+I (macOS) or Ctrl+I (Windows/Linux), type “add error handling”. You should see a diff view.
3. Autocomplete: open any code file, start typing a function signature, pause. Ghost text should appear within ~500ms. If nothing appears, check that Ollama is running and the model tag in your config matches ollama list exactly.

Common install errors and fixes

• Autocomplete outputs mangled code: FIM template mismatch. Only qwen2.5-coder, starcoder2, and deepseek-coder are safe defaults for the autocomplete role. Generic instruct models like llama3.1:8b will fail here – see the blockquote above for why.
• “Model not found”: your config.yaml model string must match ollama list output exactly, including the tag. qwen2.5-coder is not the same as qwen2.5-coder:1.5b.
• Agent mode says “does not support tools”: add capabilities: [tool_use] to the model block. Tool support varies by model and isn’t consistently documented for all Ollama models – if it still fails after adding the capability flag, the model likely doesn’t implement the required function-calling format.
• Building from source fails on sqlite3: if you clone the repo to build, you’ll likely hit prebuild-install failing on sqlite3 during node-gyp rebuild. This exact issue was reported in GitHub discussion #10484 when a user ran ./scripts/install-dependencies.sh. Fix: install the Rust toolchain (curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh), use Node 20 via nvm, then npm rebuild.
• Extension says 1.2.14: see the version trap section above.

The elephant in the room: it’s frozen

No more releases. No new bug fixes. That’s the deal – and it’s worth sitting with that for a second before committing to this toolchain.

Local features are fully self-contained: Ollama integration, config.yaml, autocomplete, chat, edit, agent mode, MCP servers. The Apache 2.0 license means the code isn’t disappearing, and forks will exist for years. Someone will maintain a community fork; several already existed before the acquisition closed.

Cloud-dependent features are another story. The Hub, Mission Control assistants, hosted authentication – these relied on Continue’s platform. According to the AI Agents Square review citing The New Stack, recurring billing was disabled after the acquisition, so paying customers won’t be charged again. The flip side: paid tier features can silently stop working without warning, and nobody’s filing bug reports anymore.

Is a frozen Apache 2.0 tool actually a problem? For local autocomplete and chat, probably not. The FIM model formats it depends on are stable, Ollama isn’t going anywhere, and the config schema is simple enough to fork if you hit a real blocker. The harder question is whether you want to invest time customizing something with no upstream – rules, prompts, MCP configs – knowing you’d have to migrate that work if you switch tools in six months.

Uninstall and cleanup

Because the project is archived, a clean uninstall is more relevant than usual – you may want to switch tools without leftover config bleeding into a new setup.

# VS Code: Extensions → Continue → uninstall
# JetBrains: Settings → Plugins → Continue → uninstall

# Remove the CLI
npm uninstall -g @continuedev/cli

# Wipe all config, sessions, logs, indexes
rm -rf ~/.continue

# Windows equivalent (PowerShell)
Remove-Item -Recurse -Force $env:USERPROFILE.continue

~/.continue holds chat history (sessions/), your workspace index, and any MCP server configs. Back it up first if you want to migrate prompts and rules to a fork later.

FAQ

Should I still install Continue in 2026 given it’s been acquired?

For solo local use with Ollama, yes – it still works and costs nothing. For a team of 50 evaluating a long-term tool, pick something with an active roadmap instead.

Can I use my GitHub Copilot subscription with Continue?

Not really. Continue supports Anthropic, OpenAI, Mistral, Ollama, Azure, Bedrock, and a long list of providers via API keys – but as of mid-2026, GitHub Copilot doesn’t expose a general-purpose completion API you can point Continue at. The typical setup is Ollama for local, or an Anthropic API key for cloud chat. If you already pay for Copilot, you’re probably better off using its own extension in parallel and leaving Continue for privacy-sensitive projects.

What’s the fastest way to test whether my machine is powerful enough?

Run ollama run qwen2.5-coder:1.5b in a terminal and ask it to write a fizzbuzz. If the response streams smoothly, autocomplete will feel fine. Then try qwen2.5-coder:7b the same way for the chat role. If the 7B model produces a full response in under 10 seconds, you’re set. If it takes 30+ seconds, drop to the 3B variant or use a cloud API for chat while keeping local autocomplete.

Next: open ~/.continue/config.yaml, paste the block above, restart VS Code, and type @codebase how does authentication work in this project in the chat panel. That single prompt is the fastest way to feel whether Continue’s context indexing is worth keeping around for you.