Theo Browne of Ping Labs spent five months and over $400,000 in OpenAI API tokens trying to build a Rust port of the TypeScript compiler. He got to 84% compatibility and stalled. He then switched to Claude Opus 5.5, which produced a working prototype in ten hours and a finished port in two weeks — for $24,047. The result, tsc-rs, is now MIT-licensed, available on npm, and types your code 11.4 times faster than TypeScript 6.
What tsc-rs Actually Is #
Before the performance claims land, the framing matters. tsc-rs is not a from-scratch TypeScript compiler written in Rust. It is a port of TSGO — Microsoft’s official Go-based rewrite shipping as TypeScript 7. Browne’s AI agents translated those Go algorithms into Rust. That distinction matters both for understanding the benchmarks and for realistic expectations about long-term maintenance.
The TypeScript compiler landscape in late 2026 looks like this:
- TypeScript 6 (March 2026): The last JavaScript-based compiler. Ships with
strictmode enabled by default, drops AMD/UMD/SystemJS support, and defaults modules toesnext. This is the baseline most teams are still on. - TypeScript 7 / TSGO : Microsoft’s official Go rewrite. Roughly 10x faster than TS6. Still in beta.
- tsc-rs : Community Rust port of TSGO. 1.61x faster than TS7, 11.4x faster than TS6. Experimental, v0.1.0, released October 7, 2026.
The Numbers #
Browne benchmarked tsc-rs across six real-world applications on an Apple M4 Pro. The geometric mean puts tsc-rs at 11.4x over TypeScript 6 and 1.61x over TypeScript 7. On the T3 Code project, tsc-rs finished in 7.25 seconds versus TypeScript 7’s 16.10 seconds. On the VS Code codebase, it completed in 4.20 seconds versus 6.84 seconds for TSGO.
One figure worth context: bun check clocks in at approximately 20.9x over TypeScript 6 — faster than tsc-rs. But bun check reports different diagnostics in some cases, which means it is not a clean drop-in for projects relying on TypeScript’s specific error output. tsc-rs passes 181,711 tests ported from Microsoft’s Go compiler and produces identical diagnostics to TSGO on TanStack Query core and Hono.
| Compiler | Speed vs TS6 | Platform | Status |
|---|---|---|---|
| TypeScript 6 | 1x | All | Stable |
| TypeScript 7 (Go) | ~10x | All | Beta |
| tsc-rs | 11.4x | Linux x64, macOS ARM | Experimental |
| bun check | ~20.9x | All | Stable |
How to Install and Use tsc-rs #
The install is one command:
npm install -D tsc-rs
Run it the same way you would run tsc:
npx tsc-rs -p tsconfig.json
No tsconfig changes required. It accepts the same flags as tsc, uses your existing tsconfig.json, and produces the same output format. It is pinned to TypeScript 7.1.0-dev and ships with Effect diagnostics built in. Current platform support: Linux x64 and macOS arm64. Windows and Linux ARM64 are not yet supported.
The AI Development Cost Story #
Browne’s announcement on X describes five months of work using GPT-5.6 Sol and GPT-6 Astra. The models generated substantial code but plateaued near 84% test compatibility and could not break through. After burning through more than $400,000 in API costs, he switched to Claude Opus 5.5.
Claude produced a v0 prototype in roughly ten hours. The full port — 181,711 tests passing — took two more weeks and cost around $24,047 total. That is a 17x cost difference in favor of Claude, though the honest framing is more nuanced: the OpenAI attempts spent months doing groundwork that Claude inherited. The $24k figure does not include the prior $400k of failed runs. Still, the end result is real: a working compiler that Browne says he could not have shipped with the other models regardless of cost or time.
The Limitations You Need to Know #
Browne is upfront that he has never read the code — every line was written by AI agents. This works because compiler correctness is well-defined by tests. But it raises legitimate questions about maintenance: TypeScript ships updates roughly every six weeks, each potentially requiring corresponding changes to tsc-rs.
Known issues in v0.1.0:
- Memory growth during extended editor sessions
- TS6059 errors in some monorepos
tsc -bmay read stale output when project references are missing- Version mismatch between README’s listed upstream revision and the tooling manifest
- No Windows or Linux ARM64 support
A Windows fork already exists, which suggests community momentum. Whether that momentum translates to sustained maintenance is still an open question.
Should You Switch? #
If you are on Linux or Apple Silicon and want faster type-checking today, tsc-rs is worth a try. It is a drop-in replacement, the install is one command, and the speed improvement is real and substantial on large projects. Treat it as experimental: keep TypeScript 7 as a fallback, validate your project’s diagnostic output, and avoid depending on it in production CI until v0.2 demonstrates a stronger monorepo and stability story.
If you are on Windows, ARM Linux, or running a large monorepo — wait. The potential is clear, but v0.1.0 is not production-ready for those scenarios. An 11.4x compilation speedup would meaningfully change the developer experience on large TypeScript codebases, and that is worth watching closely as the project matures.
The larger point tsc-rs illustrates: AI-assisted large-scale code migrations are no longer theoretical. GitHub did it with the Copilot runtime in Rust last month — 800,000 lines in 14.5 weeks. Browne did it with the TypeScript compiler. The bottleneck is shifting from “can agents write this code” to “who is responsible for maintaining what agents build.” tsc-rs is an early, honest look at both sides of that question.