Outlier vs Continue.dev: local Mac AI vs an open-source coding agent

I'm Matt Kerr and I build Outlier, so read this as a comparison written by an interested party. These two solve overlapping problems from opposite ends: Continue is a client you bring a model to; Outlier is a model you bring nothing to.

What Continue is, and what happened to it

Continue spent three years as the answer to "I want a coding assistant that doesn't lock me to one model vendor." It's a client, not a model: a VS Code extension, a JetBrains plugin, and a CLI on npm as @continuedev/cli. Continue supplies the chat panel, the edit-and-apply plumbing and the agent loop; you supply the weights — Ollama on localhost, or any cloud API you hold a key for. The license is Apache-2.0.

Then Cursor acquired the company. The homepage now reads "Continue has joined Cursor," and the README calls the repository "no longer actively maintained" and "read-only for all users" after a final 2.0.0 that removed anonymous telemetry, pulled out authentication and fixed bugs. VS Code v2.0.0 and v2.1.0 both landed 2026-06-19; the CLI ended at 1.5.47; JetBrains last shipped v1.0.67 on 2026-03-27. hub.continue.dev no longer resolves in DNS. The code survives: the repo isn't archived, and still shows ~35,000 stars and 5,100 forks.

Where each one lives, and which model does the work

Continue meets you inside the tool you already use, its best design decision — but it needs VS Code, a JetBrains IDE or Node. No native desktop app, no mobile app, and since the hub shut down, no web UI. Which model powers it is whatever you configured; Continue's Ollama guide points at open-weight options like Qwen2.5-Coder, DeepSeek-Coder 6.7B and DeepSeek-R1 32B, budgeting about 32 GB of RAM for a 32B.

Outlier goes the other way: one macOS app, Apple Silicon only (M1 or later, macOS 12+), models included. Seven tiers, from Nano 4B (2.37 GB, 6 GB RAM, 71.7 tok/s on my M1 Ultra) through Code 27B (15.13 GB, 24 GB RAM) to Plus 397B-a17b, which streams MoE experts from SSD. Six tiers are Apache 2.0 weights; Quick is Gemma Terms of Use. No extension, no plugin, no Windows or Linux build — if you work in an IDE on Linux, Continue wins by default.

Does your code leave the machine — and what does it cost?

Continue's honest answer is "up to you." Pointed at Ollama the endpoint is http://localhost:11434 and nothing leaves; the docs call that privacy-focused offline assistance. Pointed at a cloud provider, your prompts and code go to that third party on their terms. There's an official no-internet guide: disable Allow Anonymous Telemetry (which otherwise posts to PostHog), configure a local model, install from a .vsix. A real air-gapped story, better documented than most commercial assistants manage — but nothing is local by default.

One gap worth naming: the privacy policy, updated 2026-02-05, says you aren't required to provide personal data to use the open-source software — but it never states whether user code is used for training, and no opt-out is documented. It's silent, and I won't turn silence into an answer.

On price, the extension and CLI were always free; the paid tiers covered the hosted hub and credits, not the tooling. Snapshotted 2026-06-03: Starter at $3 per million tokens in and out; Team at $20 per seat per month with $10 of credits and SSO; Company at custom pricing with SAML/OIDC, BYOK and an SLA. All historical — the hub is gone. Continue now costs your provider's rates, or nothing on local weights. Outlier is free on Nano and Lite; Pro is $20/month or $149/year for all seven tiers.

Side by side

Agentic ability, repo context, and my honest numbers

On a blind slice of SWE-bench Verified — real repositories, real issues, an agent that reads code, edits files and runs tests — Outlier's local 27B measured about 45%, or 18 of 40. Code 27B scores 0.866 on HumanEval. I measured both on my own hardware; they aren't leaderboard placements; read them as a floor for what a 15 GB model on a laptop can do.

They also sit below the strongest cloud coding agents, and I won't dress that up. Frontier hosted tools resolve a higher share of SWE-bench Verified, carry far larger context windows, and don't slow down because your machine is compiling something else. On sprawling multi-repo refactors a cloud agent does better — that's arithmetic, not modesty.

Continue has no number of its own, because it ships no model. Its ceiling is whatever you attach: a frontier API outperforms anything I run locally; a 7B model won't. Repo context is the same — Continue indexes your workspace, but the window belongs to your backend, while Outlier's is bounded by tier and Mac RAM.

Where Continue.dev genuinely wins

• It's actually open source. Apache-2.0 — fork it, relicense it, ship it. Outlier's app isn't; the weights are published, which isn't the same thing. If auditability is a hard requirement, that settles it.
• Model agnosticism. Qwen2.5-Coder locally today, a frontier API tomorrow, one config change. I can't offer that.
• It's in your editor. Your selection, your diffs, your terminal, no window switching. Outlier is a separate app you paste into.
• It runs where Outlier can't: Windows, Linux, Intel Macs.
• Documented air-gapped operation, down to the .vsix install. Few commercial assistants publish that.
• No markup on inference. Local costs nothing; cloud costs provider rates.
• Three years of production use — roughly 3.72M VS Code installs and 1.18M JetBrains downloads. And the final 2.0.0, stripped of telemetry and auth, is a clean fork base.

Two asterisks: the VS Code Marketplace average is a modest 3.31 across 176 ratings, and the docs still present Continue as actively maintained, contradicting the homepage and README.

Who should pick which

Pick Continue if you want a client you can read and fork, you're on Windows, Linux or an Intel Mac, you live in VS Code or JetBrains, or you want to route hard problems to a frontier API and routine ones to local weights. Go in clear-eyed: roughly 935 open issues, no upstream security patches, and nobody shipping fixes as model APIs drift. Apache-2.0 means a maintained fork can exist — but that's now a problem with your name on it.

Pick Outlier if you're on an Apple Silicon Mac and want local inference to arrive as a product: one download, models included, no config file, no account, no caps. It's the better answer when confidentiality is the constraint — client code, regulated work, anything under NDA — because on-device is the default, not a setting you can get wrong. Know what it lacks: no web search, no image generation, no mobile app, no voice mode, no Windows or Linux, no cloud sync.

Pick a cloud coding agent if your bottleneck is raw capability on large, gnarly codebases. Plenty of people run both: local for the sensitive and routine, cloud for the hard tenth.