Update — July 21, 2026: jcode is another Rust-native coding agent harness — author benchmarks ~117 MB for ten sessions vs ~2.3 GB for Claude Code, plus native swarm and semantic memory graph. Different product (community MIT vs Anthropic bundle), same "harness performance matters" thread.
On July 19, 2026, Simon Willison published Claude Code now uses Bun written in Rust — a verification post for a swap Jarred Sumner had already described in Rewriting Bun in Rust (July 8, 2026): Claude Code v2.1.181+ (from June 17, 2026) quietly embeds Bun v1.4.0's Rust port, with ~10% faster Linux startup and little else users were meant to notice.
Willison's contribution is proof, not announcement. While Fireship's July 15 Code Report framed the 64-agent, 11-day, ~$165k mechanical migration as history's most controversial rewrite, Willison showed builders how to forensically confirm the runtime sitting inside ~/.local/bin/claude. The Hacker News thread hit 421 points — less about Claude Code than about Zig vs Rust, unsafe blocks, Anthropic's Bun acquisition, and whether a React/JS TUI shipping a Rust JS runtime is irony or dogfooding.
TL;DR — what people are asking
| Question | Direct answer |
|---|---|
| When did CC switch? | v2.1.181+, June 17, 2026 onward — stealth before loud Bun v1.4 marketing (Sumner postmortem). |
| How to verify? | strings ~/.local/bin/claude | grep 'Bun v1.4.0' — Willison Jul 19, 2026; also .rs paths in binary. |
| Public Rust Bun? | bun upgrade --canary for standalone installs — not required for Claude Code's bundled runtime. |
| User-visible change? | ~10% Linux startup per Sumner; otherwise same CLI behavior if you're updated. |
| Relation to AI rewrite story? | Same port — see Bun Zig→Rust harness deep dive for worktrees and adversarial review. |
| HN fighting about? | Unsafe Rust, 1M-line PR, vibe-coded rewrite, TUI stack — mostly culture, some valid supply-chain questions. |
What Simon Willison verified — step by step
Willison's July 19 post is short and reproducible. explainx.ai restates the checks so you do not have to infer from screenshots:
1. Strings grep on the Claude Code binary
strings ~/.local/bin/claude | grep 'Bun v1.4.0'
On a current install after updating Claude Code, Willison reports a match — evidence the embedded runtime identifies as Bun v1.4.0, the Rust port channel Sumner describes post-merge.
2. Rust source path fingerprints
Willison notes .rs file path strings inside the Claude Code binary consistent with Bun's Rust tree — not proof of build reproducibility for auditors, but stronger than assuming the old Zig runtime still ships.
3. Update Claude Code first
If grep returns nothing, you may be on pre-2.1.181 builds. Update the CLI through Anthropic's normal install path, then re-run. Version pinning matters for compliance teams documenting toolchain provenance.
4. Standalone Bun canary (optional)
For projects using Bun outside Claude Code:
bun upgrade --canary
Sumner's postmortem positions canary as the public path to Rust Bun while idiomatic cleanup continues. Claude Code bundles its runtime — this command is for your apps, not for patching Anthropic's binary.
What changed for Claude Code users — and what did not
| Area | Changed? | Notes |
|---|---|---|
| Linux startup | Yes — modest | ~10% faster per Sumner |
| Commands / flags | No intentional change | Stealth swap goal |
| TUI stack | No | Still the JS/React CLI HN loves to critique |
| Agent harness | Indirect | Faster cold starts for dynamic workflows |
| Memory leaks in long Bun sessions | Improved in Bun itself | Matters if CC plugins hold Bun workers open |
The product story Anthropic likely wants: dogfood the Rust port at scale before telling every Bun user to migrate. That mirrors parallel programming discipline — synchronize on tests and canary before declaring done.
Context from the Bun postmortem — why Rust, why stealth
Sumner's engineering case (expanded in explainx.ai's Fireship companion):
- JavaScriptCore + manual memory hybrid produced UAF, double-free, and leak classes Zig did not automatically prevent.
- ~64 parallel Claudes, PORTING.md, four worktrees, TypeScript conformance tests (~1.3M
expect()calls) — harness-first migration. - 128 bugs fixed vs v1.3.14; 19 regressions found and fixed — honest burndown.
Stealth in Claude Code was the merge gate for user trust: if a million-assertion suite stays green, runtime language is an implementation detail — until Willison teaches everyone to grep their binary.
What HN argued at 421 points — honestly
Willison's post became a lightning rod. explainx.ai separates valid from performed:
Zig vs Rust (again)
Andrew Kelley's July 9 response — engineering practice vs language — resurfaced. Neither side "won" on HN; readers split on whether $165k bought safety or faithful unsafe slop.
Unsafe Rust
Sumner discloses heavy unsafe for JSC interop. Compliance readers should treat Claude Code as shipping Anthropic-built artifacts, not auditable source you compile yourself. Willison's strings check is version forensics, not formal verification.
"Vibe-coded 1M line PR"
Mechanical port ≠ greenfield rewrite, but review surface is still enormous. Sumner's answer — adversarial reviewers, fix the workflow not the million lines — is the same lesson as risk-based human review.
TUI written in JS, runtime in Rust
Irony posts dominated X. Substantive angle: UI stack choice is independent of runtime correctness — but startup and memory bind them in one binary. Claude Code vs OpenCode token overhead readers care about what ships, not flame wars.
Anthropic acquisition
Bun inside Claude Code is S-1-grade dogfooding narrative — explainx.ai does not translate into investment advice. It is evidence Anthropic bets on agent-written systems code with conformance tests as gate.
Comparison — this post vs the Fireship rewrite coverage
| Angle | Bun Zig→Rust Fireship post | This post |
|---|---|---|
| Trigger | Fireship Code Report Jul 15 | Willison verification Jul 19 |
| Audience | Harness builders, migration architects | Claude Code users confirming binary |
| Hero command | PORTING.md / worktrees | strings … | grep 'Bun v1.4.0' |
| Depth on agents | Full 64-agent topology | Light — links outward |
Read both if you came from HN via different tabs.
Action checklist for teams
- Record CLI version in internal runbooks —
claude --versionafter each upgrade. - Run Willison's strings check on CI images that bake Claude Code into VPS workflows.
- Do not assume Zig Bun in security questionnaires — update SBOM language to Bun v1.4 Rust canary lineage where applicable.
- Separate concerns — app projects on standalone Bun need their own
bun upgrade --canarypolicy; Claude Code tracks Anthropic's bundle.
Summary
July 19, 2026: Simon Willison showed how to verify that Claude Code embeds Rust Bun v1.4.0 — strings grep, .rs fingerprints, and bun upgrade --canary for public Rust Bun. The swap shipped stealthily from June 17, 2026 (v2.1.181+) with ~10% faster Linux startup, per Jarred Sumner.
explainx.ai's read: Willison's post is the userland proof layer on top of the AI-orchestrated port. HN's 421-point debate is mostly culture; your actionable move is confirm your binary and update compliance docs — not relitigate Zig on Twitter.
Related on explainx.ai
- Bun Zig→Rust AI rewrite — Fireship and harness patterns
- Claude Code dynamic workflows
- What is an agent harness?
- Should developers stop reviewing AI-generated code?
- Claude Code VPS production workflow
- Claude Code vs OpenCode token overhead
- jcode Rust harness — swarm, memory graph, multi-session RAM (Jul 21)
- Zen of parallel programming — sync before scaling agents
- explainx.ai team skills — security-first distribution
Primary sources: Simon Willison — Claude Code uses Bun written in Rust, Jul 19, 2026 · Jarred Sumner — Rewriting Bun in Rust, Jul 8, 2026 · Andrew Kelley response, Jul 9, 2026
Verification commands, version strings, and canary channels reflect public posts through July 20, 2026. Claude Code and Bun release cadence may move faster — re-run checks after upgrades.