A Rust rewrite of the TypeScript 7 compiler had 3,003 stars in a GitHub API snapshot taken on November 8, 2026, yet no stable version. That gap is what greets anyone landing on oxc-project/tsrs. The README calls it a drop-in replacement for the TypeScript 7 compiler and language server, written in Rust, then says the team aims to release a stable version against TypeScript 7.1 with 100% conformance; aim means a target, not a delivered result. Stars, language, license, creation and push dates are all checkable in the API, but the answer to "when is it stable" is still beta and prerelease. This piece takes one repository apart: where the stars came from, what holds the stable release back, who measured each speedup figure, and whether it belongs in a production pipeline as of November 8, 2026.
Set out the batch's division of labour so nothing gets crossed. This article is a deep dive into one repository, oxc-project/tsrs, on repository facts, beta status, the staffing reality behind closed pull requests, and the provenance of three sets of speedup figures; no multi-project ranking. The OpenAI Images 2.6 GIF and SVG increments cover one product's release and official capability wording, the Cosmos 3 Nano four-path comparison explains a different set of objects, and the Google ADK 2.14.0 upgrade SOP walks through breaking changes; all different threads. Among Rust projects, the craft-anything clean-room engine is a game-engine rewrite with a dual license, while the five-tool AI metadata comparison is a parameter table for several unrelated repositories; this piece lists no checklist and compares no metadata. License-clause comparisons sit in the open-weight license comparison; Apache-2.0 is stated here as one fact, without clause-by-clause contrast.
Pin Down the Repository Facts: Which Moment Do 3,003 Stars Belong To
The GitHub REST endpoint /repos/oxc-project/tsrs returned a full snapshot at 2026-11-08T11:51:54Z: 3003 stars, Rust, Apache-2.0, created 2026-10-01T15:14:18Z, pushed 2026-11-07T17:19:37Z, forks 61, open_issues 53, default_branch main, description "Rust port of the TypeScript 7 type checker". Every figure is tied to that fetch time: an earlier probe at 11:37:30Z recorded 3002 stars, so a figure from another moment cannot overwrite this one. The owner is oxc-project; paths guessed under other organisation names returned 404. One plain address is enough for cloning: https://github.com/oxc-project/tsrs.
License agrees in three places: API field, README badge and LICENSE text all read Apache-2.0. The README adds that it is a derivative of TypeScript, copyright Microsoft Corporation, bundled lib.*.d.ts files come from TypeScript, and Not affiliated with or endorsed by Microsoft. Those lines are reproduced as written; the LICENSE text governs, with no clause reading or license comparison here.
Three Sentences in the README: Position, Target, Staffing
Three README sentences deserve to be read word for word. First: "tsrs is a drop-in replacement for the TypeScript 7 compiler and language server, written in Rust." That is the position, a direct replacement in Rust. Second: "We aim to release a stable version against TypeScript 7.1 with 100% conformance." That is the target, and aim is a promise not yet redeemed, not an acceptance result. Third: "Pull requests remain closed because we don't have the bandwidth to review them." That is the reality promotional copy most easily swallows: PRs stay closed for lack of review staffing. Issues remain open, but merges are not driven from outside.
The VoidZero October review, published November 5, 2026, adds "tsrs is not a fork", close work with the TypeScript team, a 100% conformance goal for TypeScript 7.1, and performance improvements fed back upstream. That is outside-repository wording, attributed as such, not a README statement. Read together, the project describes itself with restraint: a Rust replacement, an unmet conformance target, a community short of hands. 3,003 stars measure attention, not whether the work is done.
The beta Cadence: What docs/STATUS.md Records
docs/STATUS.md runs to 1,041 lines. Version strings include 0.23.0-ts7.1.0-beta and 0.22.0-ts7.1.0-beta, beside a 0.20.0 beta heading; the body uses "creates a GitHub prerelease", "npm's beta tag is assigned manually" and "release candidate". Read together, releases generate a GitHub prerelease, the npm beta tag is assigned by hand, and the wording has reached a release candidate, yet nothing announces that stable has landed. The accurate description is beta, prerelease, stable version not released: the label stops at prerelease.
One cadence detail matters. The README says --fast and its default preset are available in 0.20.0 and later, also writable as TSRS_FAST=1. The default is four checkers capped by file count, following Go's file assignment and checker history; fast adapts the checker count, groups tasks, adds caching, and preserves strict, skipLibCheck and noCheck. This is why running the benchmark with and without --fast lands in different ranges, and why one result under one version and one switch is not the whole picture.
Three Sets of Speedup Numbers, and Why They Do Not Confirm Each Other
More than one set of "N times faster" figures circulates under this name, from different publishers, against different baselines, on different machines, with different methods. They cannot be converted, added, or used to explain one another. Each is set out below with its comparison target attached.
Outside the repository, the Cloudflare official blog states "up to 8x faster" and "~70% less memory", against TypeScript 7.0 with the same number of checker workers. It is not in any benchmark file in the repository; quote it together with "against TypeScript 7.0", not as a reproducible in-repo measurement.
Inside the repository, bench/RESULTS.md, fetched at 2026-11-08T11:51:54Z at 8,295 characters, is the project's own reproducible benchmark: 19 projects, an Apple M3 Max with 14 cores and 36GB, measured 2026-10-28, medians of ten interleaved runs with Hyperfine 2.0.0, five for TypeScript, comparing TypeScript 7.1.0-beta with tsrs 0.20.1-ts7.1.0-beta, commands carrying --noEmit --incremental false --pretty false. Result: wall-clock 1.71x-6.62x faster, peak memory 1.70x-2.75x less, and TypeScript, tsrs and tsrs --fast report identical diagnostics, file, position and code, for every project. bench/README.md says projects.json fixes repositories and commits, bench/publication.json holds full-precision values, and charts and report reproduce byte for byte. The classification matters: a first-party self-test, not independent third-party measurement; no third-party retest appears in the fetched README, RESULTS.md, bench/README.md or STATUS.md.
The third-party set stays separate. ImportStatic reported about 2.7x against TypeScript 7.1 beta using TypeScript's own source; Frank Chu reported 5.3x and memory -72% under 400 files, 2 cores, default settings, against TypeScript 7.0.2; Cloudflare's own chart lists four projects at 3.8x-8.1x. Daniel Rosenwasser of Microsoft responded that the gains do not all come from the language and that part of the TypeScript 7.1 improvement comes from tsrs feeding upstream. These are observations under their own conditions, not rows for one merged table.
Collapsing the three into "eight times officially, 2.7 times measured" is the common misreading. Baselines already differ: TypeScript 7.0, 7.1.0-beta, 7.0.2. Machines, project sets, whether --fast was on, run counts behind the median, and publishers differ too. The numbers can be neither converted nor used to explain one another; each answers only its own conditions. Putting multipliers from different benchmarks side by side in a headline implies one side exaggerated, which this material does not support. What can be written: three sets each stand alone and do not corroborate each other. What cannot: one unified multiplier folded out of them.
Capability Boundaries: WASM, the Language Server and Type-Aware Lint
The README claims a language server, and the repository carries docs/LSP.md, docs/USAGE.md, docs/NODE_API.md, docs/COMPATIBILITY.md, docs/DEFAULTS.md and docs/STATUS.md. @tsrs/wasm targets wasm32-wasip1, WASI Preview 1, letting type checking run in browsers, Workers and embedded environments; a tsrs playground allows checking and code emission in a browser. The VoidZero October review mentions type-aware lint in Oxlint, 61 rules, and the playground as planned directions from outside the repository, not delivered capabilities.
Per the README, install is two lines:
pnpm add -D tsrs
pnpm exec tsrs --noEmit --fast
TSRS_FAST=1 pnpm exec tsrs --noEmitOutside material, the Cloudflare blog, InfoQ and VoidZero, gives npm install -D tsrs with npx tsrs -p . --noEmit, and vp add -D tsrs with vpx tsrs --noEmit --fast; entry points and moments differ, so they should not be welded together. A hands-on developer test reports tsrs exits 2 on a type error while tsc exits 1; the README does not mention it, so verify before wiring any pipeline.
How to Read Claims Like "a Compiler Written Overnight by an Agent"
More eye-catching claims live outside the repository and need their sources marked. "An AI agent wrote a compiler overnight" comes from the Cloudflare blog, InfoQ and Owner.com's Schwenk: according to Schwenk's own account, he wrote no code, and Claude Opus 5.5 plus more than twenty helper agents built it in one night against TypeScript's official 15,197 conformance tests. Owner.com separately says a 37,000-file back-end type check fell from 80 seconds single-threaded to 7 seconds in a pipeline, a company self-report. "Cloudflare and VoidZero collaborating" and incubation in the Oxc toolchain are likewise outside promotional framing. These can be written only as "according to reports" or "according to Schwenk's own account", never as official positioning or a headline modifier; the README keeps that replacement sentence as its self-description. The announcement date, October 20, 2026, comes from a flaviocopes record outside the repository, the VoidZero October review is dated November 5, 2026, and the API created field is 2026-10-01, so as of the November 8, 2026 fetch it had existed since October 1; a coverage date is not a creation date.
Can It Replace Production as of November 8, 2026
Pull the above together and the judgment is not hard. First, the stable version has not been released; 100% conformance is an aim, not an acceptance result. Second, docs/STATUS.md shows a hand-assigned beta tag and version strings still carrying beta, so the cadence is prerelease. Third, PRs remain closed for lack of review staffing, so external fixes are not community-driven. Fourth, the three speedup sets do not corroborate each other, so none proves it is fast enough to switch. Fifth, details like the exit-code difference need checking oneself. The reasonable use is a side install, running --noEmit on your own project, comparing diagnostics and timing, and filing issues on performance, memory and conformance, which is what the README invites. The poor fit is writing it into production dependencies or a delivery pipeline, let alone declaring on its behalf that it can stand in for tsc. Replacement becomes defensible only once a stable version lands and 100% conformance has a checkable acceptance result. 3,003 stars speak to attention, not readiness; a repository that keeps PRs closed, states its target clearly and separates benchmarks by source is better read on the facts than on handsome multipliers.
FAQ
Q1: What exactly is oxc-project/tsrs?
A1: A replacement implementation of the TypeScript 7 compiler and language server written in Rust, described verbatim in the README as a drop-in replacement for the TypeScript 7 compiler and language server, written in Rust. Its exact GitHub path is oxc-project/tsrs; the API snapshot at 2026-11-08T11:51:54Z gives 3003 stars, Rust and Apache-2.0. It is beta and prerelease, with the stable version not released.
Q2: Do 3,003 stars mean it is stable and production-ready?
A2: No. 3,003 is a snapshot from 2026-11-08T11:51:54Z; 100% conformance uses the word aim, and docs/STATUS.md still shows 0.23.0-ts7.1.0-beta, a GitHub prerelease and a manually assigned npm beta tag, with no record of stable landing. PRs remain closed for lack of review staffing. Stars measure attention, not readiness.
Q3: Why are there several multipliers, 8x, 2.7x and 5.3x?
A3: Three unrelated sets: the Cloudflare official blog's up to 8x faster and ~70% less memory against TypeScript 7.0, outside-repository wording; the project's own bench/RESULTS.md self-test with 19 projects, an Apple M3 Max, 2026-10-28, Hyperfine medians, against 7.1.0-beta, giving 1.71x-6.62x faster and 1.70x-2.75x less; and third-party ImportStatic at about 2.7x with Frank Chu at 5.3x / -72%, each with its own machine and baseline. Baselines, machines, methods and publishers differ, so they cannot be converted, added, or used to explain one another.
Q4: Is "an AI agent wrote a compiler overnight" official positioning?
A4: No. It comes from the Cloudflare blog, InfoQ and Owner.com's Schwenk, whose account describes Claude Opus 5.5 plus more than twenty helper agents building it against TypeScript's official 15,197 conformance tests. It is outside promotional or self-reported framing, marked "according to reports" or "according to Schwenk's own account". The README's official positioning is the drop-in replacement sentence, with no such headline modifier.
Q5: Which commands should I start with?
A5: Per the README, pnpm add -D tsrs, then pnpm exec tsrs --noEmit --fast, equivalent to TSRS_FAST=1 pnpm exec tsrs --noEmit. Outside material gives npm install -D tsrs with npx tsrs -p . --noEmit and vp add -D tsrs with vpx tsrs --noEmit --fast; entry points differ, so do not mix them. Compare diagnostics and timing on a side project before any delivery pipeline, and verify the exit-code difference before CI, since tsrs exits 2 on a type error while tsc exits 1.
Related reading