Continue.dev

Continue.dev is the *bring-your-own-model coding assistant* — an open-source IDE extension that gives you Copilot-style chat, inline edits, and autocomplete while letting you point it at a fully local model, so your source code never has to leave your machine.

Background. Continue launched in 2023 as an *Apache-2.0* extension for VS Code and JetBrains. Its defining choice is model-agnosticism: instead of hardwiring one vendor's cloud, you configure the provider — *local Ollama / LM Studio* for full privacy, or OpenAI / Anthropic / OpenRouter / others via API key when you want frontier capability. That open license plus the local-model path is why it earns an A in /ai: it's the category's credible privacy-respecting alternative to closed, telemetry-heavy assistants.

What you trust. The trust model is "you choose where the data goes." *Open source* means the extension's behaviour — what it sends, when, to which endpoint — is auditable rather than asserted. Configure a *local model* (Ollama/LM Studio) and your prompts, code context, and completions stay entirely on your hardware; nothing transits a third party at all. Use a hosted API and you're trusting that provider, but Continue itself isn't an intermediary harvesting your code. There's *no mandatory account* tying your editing to an identity. Compared to mainstream AI-in-the-editor — which ships your code and prompts to a vendor by default — this is a structurally different posture: the tool is a conduit you control, not a data funnel.

Operational specs. It runs as a *VS Code / JetBrains extension* and provides chat, *inline edit* (select code, describe the change), autocomplete, and *codebase indexing* (local embeddings for "ask my repo" retrieval). Everything is driven by a config file where you declare your model(s), so swapping between a local Llama/Qwen/DeepSeek and a hosted model is a one-line change. With a local stack (Ollama + an open-weight code model) the entire loop — index, retrieve, prompt, complete — happens offline. Free and open-source; you pay only for whatever hosted model you opt into, if any.

Philosophy. Code is among the most sensitive data a developer handles — proprietary logic, secrets, unreleased products — yet the dominant AI assistants stream it to a vendor as the price of admission. Continue's thesis is that you shouldn't have to choose between AI assistance and confidentiality: keep the assistant, control the model. Open-sourcing it and supporting local inference are the two moves that make that real — you can read what it does, and you can run it where nothing leaves your laptop.

Grade rationale. A in /ai. The grade reflects an open (Apache-2.0) codebase, genuine local-model support (full offline privacy), no mandatory account, multi-IDE reach, and a config-driven design that doesn't lock you to one provider. It's the privacy-conscious default for AI in the editor. The caveats are about the local-model tradeoff and the BYO-key paths, not the tool's integrity.

Useful when. Use Continue when you want AI help *inside your editor* but your code is confidential, under NDA, or simply something you'd rather not feed a cloud vendor — pair it with a local Ollama model for an offline assistant. It's also the right pick when you want to switch freely between local and hosted models per task (cheap/private local for routine edits, a frontier API for hard problems) without changing tools.

Caveats. Local models are only as good as your hardware — a laptop-sized model won't match a frontier cloud model's reasoning, so the privacy/capability tradeoff is real and you should size expectations to your GPU. If you configure a *hosted* API provider, your code and prompts go to *that* vendor under their policy — Continue's privacy guarantee is only as strong as the backend you point it at, so the local path is the one that's truly private. As open-source dev tooling it expects some setup (installing a model runner, writing config). And codebase indexing builds local embeddings that live on disk — fine, but be aware of where. None of this dents the A: the ability to run a competent assistant with zero data egress is exactly what the category lacks elsewhere.