Outlier vs Sourcegraph Cody: local coding AI vs enterprise cloud AI

Disclosure first: I'm Matt Kerr, and I build Outlier, so read this as a comparison written by an interested party. Every Cody claim comes from Sourcegraph's own docs, linked at the bottom; every Outlier number is one I measured.

What Cody is in mid-2026

Sourcegraph's docs describe Cody as an AI coding assistant that uses recent LLMs plus your development context to help you understand, write and fix code faster. The page carries a badge: "Supported on Sourcegraph Enterprise." That badge is the whole story.

On June 25, 2025, Sourcegraph announced that Cody Free and Cody Pro would shut down on July 23, 2025, and that Enterprise Starter would stop including Cody. Departing users were pointed at Amp with $10 and $40 in credits; Enterprise customers were unaffected, and Sourcegraph said it remains fully committed to that offering. Both install guides now list the same gating prerequisite: a Sourcegraph Enterprise account with Cody enabled.

The clients still ship — VS Code v1.155.0 landed 2026-06-17 with ~868,000 installs and a 3.98/5 average. Weigh that against the quiet: Cody is absent from Sourcegraph's homepage, its pricing page and the entire 2026 changelog, and goes unmentioned in the December 2, 2025 post announcing that Sourcegraph and Amp are splitting into separate companies.

Where each one lives

Cody is an editor extension first. Documented clients: VS Code, JetBrains IDEs, Visual Studio (marked Experimental), the Sourcegraph web app, and a CLI installed with npm install -g @sourcegraph/cody on Node.js v20+ — itself "in the Experimental stage for Enterprise accounts." No standalone desktop app and no mobile client are documented.

Outlier is the opposite shape: a native macOS app for Apple Silicon, M1 or later, macOS 12+. No Intel build, no Windows, no Linux, no mobile app, no editor extension. Its coding agent runs inside its own window rather than inside your editor: point it at a project and it reads the files, edits them, and runs your tests. What it can't do is type alongside you — want ghost text in your gutter as you type? Cody does that; Outlier doesn't.

Which model runs, and where your code goes

Cody routes to frontier cloud models: Claude Opus 4.8 through 4.5, Sonnet 5 and 4.5, Haiku 4.5; Gemini 2.5 Pro and 3.5 Flash; GPT-5.4 down through GPT-4o and o3. Chat defaults to Sonnet 4.5. Autocomplete is separate — DeepSeek V2 Lite Base and an Autoedits fine-tune of DeepSeek Coder V2 via Fireworks.ai: open weights, consumed as a hosted API rather than run by you.

Inference is entirely remote. The FAQ says Cody sends code snippets of up to 28 KB per request to a third-party cloud LLM, and that even a self-hosted instance needs internet access for those calls, routed by default through Cody Gateway. The written guarantees are strong: Sourcegraph and its partner LLMs "do not use code from Cody Enterprise or Cody Pro teams to train models," partner LLMs retain nothing "beyond the time it takes to generate the output," fine-tuning is opt-in, and enterprises can bypass the Gateway entirely. One caveat — the only documented local-inference path is the Ollama page, which still says support is "available for Cody Free and Pro plans," switched off in July 2025.

Outlier's answer is structural rather than contractual. Seven tiers download once and run locally, from Nano 4B (2.37 GB) to Plus 397B-a17b (209 GB, streaming experts from SSD at ~11 GB peak RSS). The coding tier is Code 27B: 15.13 GB, 24 GB of RAM, weights published at huggingface.co/Outlier-Ai. No account, no cap, Wi-Fi off if you like — there's no data policy to read because there's no transfer.

Cost, and the side-by-side

Cody has no self-serve purchase and no published per-seat price. It appears nowhere on sourcegraph.com/pricing, where the only figure is an Enterprise plan "Starting at $16K." Outlier is free for Nano and Lite, a one-time $249 for all seven tiers, Founders Lifetime at $249. These aren't competing price tags: one buys a platform, the other licenses a Mac.

The coding receipts, honestly

Here's my side, measured rather than lifted off a leaderboard. On a blind slice of SWE-bench Verified, the local 27B resolved about 45% — 18 of 40. Code 27B scores 0.866 on HumanEval. Core 27B matched Claude Opus on 98.9% of rubric checks in a 54-prompt head-to-head.

Those are my own runs on my own hardware — not a submission, not audited, and 40 instances is a small slice. Forty-five percent is a real result for a 15 GB model running with the network off. It's also below what the strongest cloud coding agents resolve, and Cody hands your prompt to exactly those models. On a hard multi-file bug in an unfamiliar repo, I'd expect Opus 4.8 through Cody to land the fix more often. Sourcegraph publishes no resolve rate for Cody that I could find, so there's no number to set beside mine.

The context gap is starker than the model gap anyway. Cody builds context from keyword search, the Sourcegraph Search API against remote repositories, and the code graph, plus @-mentions of files, symbols and repos — so it answers questions about code that never touched your laptop. Outlier can't, and local model quality doesn't fix that.

Where Sourcegraph Cody genuinely wins

Cody's moat is real, and a local Mac app can't cross it.

• Organization-scale context. Retrieval spans every remote repository the company has indexed. For a 5,000-repo org, that's context Outlier can't see, because the code isn't on the machine.
• Permission-aware retrieval. The FAQ states retrieval ensures only code the user has read permission for is retrieved — a hard problem a single-user tool never has to solve, and a requirement in regulated shops.
• No hardware ceiling. Opus 4.8 and GPT-5.4 don't care about your RAM, quantization or thermals, and the model list refreshes without a re-download.
• Written data guarantees. Contractual no-training on Enterprise code, zero retention, opt-in fine-tuning, and bring-your-own-key if you'd rather the vendor never see a prompt.
• Enterprise governance. RBAC gating, admin-chosen models, Context Filters excluding whole repositories from Cody's context (regex on repo name), an org-wide Prompt Library. A desktop app has no answer.
• One assistant across a mixed toolchain — VS Code, eleven JetBrains IDEs, Visual Studio, the web — with nothing to provision locally. It behaves the same on an 8 GB laptop as on a workstation.

Who should pick which

Pick Cody if your company already runs Sourcegraph, your code lives in hundreds of repositories nobody clones locally, you need permission-scoped retrieval and admin governance, and you want frontier-model quality on hard multi-file work. One caveat: the Sourcegraph index is the differentiator, so without an indexed deployment you're paying enterprise prices for an ordinary IDE chat panel.

Pick Outlier if you're an individual or small team on Macs, your code needs to stay put by architecture rather than by contract, you work offline or on locked-down networks, or you simply can't buy Cody — which, since Free and Pro were retired, describes most developers.

My side's gaps, plainly: no image generation, no mobile app, no voice mode, no Windows or Linux build, no cloud sync. The coding tier wants 24 GB of RAM. And Outlier never sees a repository that isn't on your disk.