Verdict: If you write code comments, commit messages, documentation, or notes in privacy-sensitive environments and want instant grammar checks without shipping every keystroke to a cloud server, Harper is the best tool available in 2026. It's free, open-source, built in Rust, runs entirely on your device, and delivers suggestions in under 10 milliseconds. It will not replace Grammarly's tone rewrites or AI-powered sentence coaching — but for the surfaces Grammarly never covers (your editor, your terminal, your code), nothing else comes close.
Last verified: 2026-07-22 — Harper is at version 1.3+ on GitHub with 11,000+ stars. Grammarly Pro costs $144/year billed annually. Prices and features change — re-check before committing.
What is Harper?
Harper is a privacy-first grammar checker written in Rust. It runs entirely on your device — no text is ever sent to a server. It was created by developer Elijah Potter, who built it after years of frustration with Grammarly's cost, inconsistent suggestions, and data collection model. In November 2024, Automattic (the company behind WordPress.com) acquired Harper and brought Potter on as a Code Wrangler, with the stated goal of embedding the checker across WordPress, WooCommerce, and Jetpack. The project remains Apache-2.0 licensed and actively developed on GitHub.
What makes Harper different is its architecture: instead of a machine learning model that probabilistically flags issues and changes its mind between sessions, Harper uses deterministic, handwritten rules. The same sentence gets the same verdict every time. That sounds like a limitation until you've lived with it — predictability in a tool that fires on every keystroke is a feature, not a compromise.
How does Harper compare to Grammarly?
Grammarly is a cloud-based AI writing assistant that charges $144/year for its Pro plan and ships every word you type to AWS servers for analysis. Harper is free, open-source, and never sends your data anywhere. But the comparison isn't apples-to-apples — they serve fundamentally different surfaces.
| Feature | Harper | Grammarly Free | Grammarly Pro ($144/yr) |
|---|---|---|---|
| Price | Free | $0 | $12/month annually ($144/year) |
| Open source | Yes (Apache-2.0) | No | No |
| Runs offline | Yes, fully | No — cloud required | No — cloud required |
| Data sent to servers | None | All text processed on AWS | All text processed on AWS |
| Speed | < 10ms per check | Network latency | Network latency |
| Language | English only (5 dialects) | 6 languages | 6 languages |
| Grammar rules | ~287 handwritten Rust rules | ML model (probabilistic) | ML model (probabilistic) |
| Tone / clarity / rewrites | No | No | Yes — full sentence rewrites, tone adjustment |
| Plagiarism detection | No | No | Yes |
| Code editor support | VS Code, Neovim, Helix, Emacs, Zed | Limited (VS Code extension) | Limited |
| Browser extension | Chrome, Firefox | Chrome, Firefox, Safari | Chrome, Firefox, Safari |
| Deterministic | Yes — same input, same result | No — ML model changes verdicts | No — ML model changes verdicts |
| Desktop app | macOS (system-wide) | macOS + Windows | macOS + Windows |
The core tradeoff: Grammarly does things Harper doesn't attempt — tone rewrites, clarity coaching, plagiarism detection, multilingual support. If you need those, keep paying for Grammarly. Harper's value is in the places Grammarly will never reach: your code editor, your terminal, your commit messages, and any environment where sending your text to a third-party server is simply not acceptable.
Why does Grammarly's data collection matter?
Grammarly's own privacy policy and support FAQ confirm that when you use their products, your text is sent to their servers hosted on Amazon Web Services in US-based data centers. Their algorithms analyze your text, writing behavior, and usage data to provide suggestions. While Grammarly explicitly states they don't sell user content and that you retain all rights to your writing, the fact remains: everything you type while the extension is active is processed on their infrastructure.
Grammarly is so aware of the privacy concern that their support center has a dedicated article titled "Is Grammarly a keylogger?" in which they argue that their product is not malicious surveillance software. They make a fair distinction — Grammarly is a legitimate product with transparent terms, and they do exclude sensitive fields like passwords and credit cards. But the fundamental architecture is clear: your text leaves your machine, hits their servers, and comes back with suggestions.
Harper's architecture is the opposite: text never leaves your device. There is no server to hack, no data to leak, no retention policy to worry about. The privacy guarantee isn't a policy page — it's a property of the software itself. For a deeper look at why infrastructure architecture matters when choosing AI tools, see our analysis on LLM infrastructure security in 2026.
How fast is Harper compared to Grammarly?
Harper claims sub-10-millisecond response times per check, and the architecture backs this up — because all processing is local with no network round-trip, the only latency is your CPU. Grammarly, by contrast, requires a network round-trip for every check, which introduces latency that scales with your connection quality.
According to the Harper GitHub README, the engine takes "milliseconds to lint a document" and uses "less than 1/50th of LanguageTool's memory footprint." Automattic's acquisition announcement cited sub-20ms response times, while the project's own documentation targets sub-10ms. No independent benchmarks exist as of July 2026 — these are vendor-claimed numbers.
The practical difference is immediate: when Harper flags an error, the underline appears before you finish typing the next word. There is no spinner, no waiting for the cloud. For a tool that fires on every keystroke, this speed is what makes real-time editing tolerable.
What are Harper's limitations?
Harper is not a Grammarly replacement for everyone. Here's where the ceiling is:
English only. Harper supports five English dialects (American, British, Canadian, Australian, and Indian English). If you write in another language, Harper doesn't work. LanguageTool, by contrast, supports 30+ languages.
Rule-based, not AI-powered. Harper uses approximately 287 handwritten Rust rules (as of the video source's count — the rule count grows with each release). This means if no rule matches a particular error, nothing fires. A sentence like "What your name?" or "What day today?" will pass cleanly because they don't match any rule pattern — even though they are clearly broken. Grammarly's ML model catches these because it has statistical patterns trained on millions of texts. LanguageTool has over 6,000 rules and optionally uses n-gram data for context-aware checks, though that requires downloading ~8 GB of data and memory to match.
False positives. Harper's GitHub issues include over 240 reports mentioning false positives. The most-cited example: Harper flags correctly-spelled brand names like "Cloudflare" and "Wikimedia" for review. These are being fixed weekly, but you will encounter them.
No tone, clarity, or sentence rewrites. Harper catches grammar, spelling, punctuation, and basic style. It does not rewrite your sentences, adjust your tone, or coach your clarity. If those features are why you pay for Grammarly, Harper won't help.
No independent benchmarks. The speed and memory numbers are vendor-claimed. Some Neovim users reported the language server consuming over 1 GB of RAM. No third-party benchmark suite has been published.
Automattic governance concerns. After two years of WordPress community drama (the WP Engine legal battle and Matt Mullenweg's actions), some developers on Hacker News have stated flatly they won't adopt anything Mullenweg controls. This is a governance risk, not a technical one — but it's worth knowing before you build a workflow around a tool owned by a company in the middle of a community schism.
How to set up Harper in your workflow
Getting Harper running takes under five minutes. Here are the setup paths for the most common environments:
1. Install in VS Code (or any LSP-compatible editor)
# macOS (Homebrew)
brew install harper
# Or install via crates.io (requires Rust toolchain)
cargo install harper-ls
In VS Code, search for "Harper" in the extension marketplace and install it. For Neovim, Helix, Emacs, Zed, and Sublime Text, Harper ships as a language server (harper-ls) that drops in with one line of configuration per the Harper documentation.
2. Install the browser extension
Harper is available on the Chrome Web Store and Firefox Add-ons. The extension runs entirely in your browser using WebAssembly — no server calls.
3. Try it without installing anything
Go to writewithharper.com and start typing. The entire grammar engine runs in your browser tab via WebAssembly. This is the zero-commitment test drive: if you want to see what Harper catches without installing anything, this is the fastest path.
4. Obsidian plugin
Harper has a dedicated Obsidian plugin with over 200,000 downloads. It checks your notes inline and skips code fences — the exact thing Grammarly trips over. If you write notes in Obsidian, this is the integration that matters most.
Harper vs LanguageTool: which free alternative should you pick?
If you're choosing between the two major free/open-source grammar checkers, here's the decision framework:
| Use case | Pick | Why |
|---|---|---|
| Developer editing code comments + docs in-editor | Harper | LSP-native, sub-10ms, zero-setup offline |
| Multilingual writer (non-English) | LanguageTool | 30+ languages vs Harper's English-only |
| Self-hosted server for a team | LanguageTool | Mature self-hosted server with API + n-gram support |
| Privacy-critical air-gapped environment | Harper | No server dependency at all — single binary, runs anywhere |
| Maximum grammar coverage (most rules) | LanguageTool | 6,000+ rules + optional n-gram context vs Harper's ~287 |
| Instant feedback, no waiting | Harper | <10ms local vs LanguageTool's seconds-long linting |
| Low memory footprint | Harper | <1/50th of LanguageTool's memory |
LanguageTool (open-source, LGPL-2.1, ~14,700 GitHub stars) is the more mature project — it launched in 2005 and supports 30+ languages. But its self-hosted context checks require downloading ~8 GB of n-gram data and enough RAM to hold it. Harper's advantage is its zero-dependency, zero-server, zero-network design. You install it, it works, and nothing ever leaves your machine. This makes it a natural fit for anyone building a self-hosted AI workspace where you control the infrastructure.
What this means for you
If you're a developer, writer, or small business owner who works with sensitive content — client emails, internal docs, code reviews, legal drafts, medical notes — Harper lets you catch typos and grammar errors without the inherent risk of sending that text to a third-party server. The tools you use daily (VS Code, Neovim, Obsidian, your browser) get instant grammar checking that never touches the cloud. For a broader look at building privacy-preserving AI workflows, see our guide on how to build privacy-preserving AI agents in 2026.
If you need AI-powered tone coaching and full-sentence rewrites, keep Grammarly Pro. Harper doesn't compete there. The two tools can coexist: Grammarly in your browser for email and document drafting, Harper in your editor for code comments, commit messages, and technical docs.
The real power move is using both strategically. Use Grammarly where the cloud features matter and the content isn't sensitive. Use Harper everywhere else — where privacy is non-negotiable and speed matters more than AI sophistication.
FAQ
Q: Is Harper really completely free?
A: Yes. Harper is free in every sense — no credit card required, no freemium upsell, no subscription. The source code is freely available under the Apache-2.0 license on GitHub, and the project is backed by Automattic, which funds its development. If you're building a cost-effective AI toolkit, Harper pairs well with other free AI tools that actually deliver value in 2026.
Q: Does Harper work offline?
A: Yes. Harper runs entirely on your device with no network connection required. Once installed, you can disconnect from the internet and it will continue to check your writing in real time. This is its core differentiator from both Grammarly and LanguageTool's cloud mode.
Q: Can Harper replace Grammarly?
A: Not if you use Grammarly's Pro features — tone rewriting, clarity suggestions, plagiarism detection, or multilingual support. Harper catches spelling, grammar, and punctuation errors but does not attempt AI-powered sentence coaching. If you only use Grammarly's free tier for basic grammar checks, Harper is a viable full replacement.
Q: What editors does Harper support?
A: Harper supports VS Code, Neovim, Helix, Emacs, Zed, and Sublime Text via its language server (harper-ls). It also has browser extensions for Chrome and Firefox, an Obsidian plugin, a native macOS desktop app, and a web-based demo that runs entirely in your browser via WebAssembly.
Q: Is Harper's data really private?
A: Harper has no data to protect because it never collects any. There is no telemetry, no cloud pipeline, no account to sign up for, and no LLM in the loop. Your text is processed by a local Rust binary and never transmitted anywhere. The privacy guarantee is architectural, not a policy.
Q: Who owns Harper and is it safe to depend on?
A: Harper was created by Elijah Potter and acquired by Automattic (WordPress.com parent company) in November 2024. It is Apache-2.0 licensed, meaning even if Automattic stops supporting it, the community can fork and continue development. Some developers have raised governance concerns around Automattic CEO Matt Mullenweg's recent actions, but the open-source license provides a safety net.
Discussion
0 comments