Dev Tools AI Apps

Build hosted mini-apps by chatting — describe an app and the assistant deploys it to {slug}.skillsafe.ai.

Edit any PowerPoint with plain English - retitle, reword, restructure, reorder, and reformat slides with a live preview, then download the same deck with your edits applied. The file stays in your browser; only the slides being edited are sent to the AI model.

Paste a bug - symptoms, stack trace, the code involved, what changed recently - and get a systematic debugging plan: an honest read of reproducibility, root-cause hypotheses ranked by evidence, the cheapest experiments first, surgical instrumentation, conditional fix directions and open questions. The quick scanner runs free in your browser. Derived from @wshobson/debugging-strategies.

Paste a JavaScript or TypeScript module and get a publish-ready SDK documentation article: definition-first opening, working code against your real exports, YAML frontmatter, a style lint that hunts down AI writing tells, and honest notes naming the gaps only the author can fill. Derived from @tldraw/write-docs.

Paste raw incident notes - Slack scrollback, an on-call log, pager timestamps - and get a complete blameless postmortem: an honest severity call, a reconstructed timeline, a 5 Whys root-cause chain grounded in your notes with evidence, contributing factors, what went well and poorly, and prioritized ticket-ready action items exported as CSV. A free instant prescan finds the timestamps, blame-language and incident signals in your paste before you run anything. Derived from the @wshobson/postmortem-writing skill (MIT).

Convert Word, Excel, PowerPoint, OpenDocument, EPUB, RTF, HTML, CSV and text-layer PDFs to clean GitHub-Flavored Markdown, entirely in the browser, with an optional metered AI pass that repairs heading hierarchy, undeclared PDF tables, merged headers, reading order and OCR noise, then reconciles its output against the parser's measured facts. Modelled on anydoc, the open-source (MIT) document parser by Firecrawl (github.com/firecrawl/anydoc); reimplemented in JavaScript for this runtime. Not affiliated with Firecrawl.

Paste a git diff, commits, and your notes - get a reviewer-first pull request title and description in Conventional Commits style, with a client-side diff stats breakdown and title linter. Derived from @tldraw/write-pr.

Paste an Azure workload description - get an enterprise architecture plan: resource topology, network design, Well-Architected alignment, risks, and a Bicep or Terraform starter. Derived from @microsoft/azure-enterprise-infra-planner.

Paste frontend source — components, stylesheets, theme files or a Tailwind config — and get the project design system reverse-engineered: six codification scores, every inconsistency quoted with its concrete fix, a consolidated token block built only from your own values, and a complete ready-to-commit DESIGN.md. Free instant token scan runs in your browser. Derived from @google-labs-code/extract-design-md.

Turn a one-line objective plus your project facts into a dependency-ordered construction plan any coding agent can execute cold: self-contained context briefs, one-PR-sized steps, real verification commands, observable exit criteria, and parallel waves computed from the dependency graph. Paste an existing plan and the analyzer audits it free in your browser; forks only you can settle come back as open decisions, never guesses. Derived from @affaan-m/blueprint.

Paste a CSV or describe your data and what it should communicate. A free in-browser profiler types your columns, picks a heuristic chart and sketches it instantly; the AI then recommends the right chart form with exact encodings, the alternatives and why they lose, misleading-chart pitfalls, accessibility guidance, and runnable code for matplotlib, plotly, Vega-Lite, Recharts, D3 or ECharts. Critique mode reviews a chart you already built. Derived from @shubhamsaboo/visualization-expert.

Web performance, arithmetic first. Drop a Lighthouse JSON report, a PageSpeed Insights response, a HAR capture, a DevTools resource table or just your metric readings, and your browser does the whole measurable half for free: every Core Web Vital banded against the published thresholds, page weight by resource type in wire and decoded bytes, a performance budget derived line by line where your own target replaces the preset instead of stacking on it, a savings ledger in which each resource is claimed by exactly one rule so the totals cannot double count, repeat-visit cache savings kept in their own ledger for the same reason, twenty-three findings each reading only the fields it is about, and a generated Lighthouse budgets.json, priority snippet, three CSVs and a full report. The AI pass adds the judgement: which finding is actually holding the largest paint back, the fix written for your stack, the work ranked by impact against effort, and a verdict on every finding including the false positives. A derived work of the web-performance-optimization skill by @sickn33. Analysis of a capture you supply, not a measurement of your page.

Paste the material behind a technical decision — rough notes, a design discussion, or an existing ADR draft — and get an architect's review: a ready / needs-work / not-ready verdict, a section-by-section audit, prioritized gap findings, and a complete, committable MADR-style Architecture Decision Record with drivers, options, consequences and a confirmation plan. A free client-side prescan flags single-option decisions, missing drawbacks, weasel words and reversibility gaps as you type. Derived from @wshobson/architecture-decision-records (MIT).

Paste your system brief — what you are building, the expected load, the constraints — and get a senior backend architect review: a readiness posture, the component plan with data ownership, prioritized findings with concrete design fixes, quick wins and focus areas. A free instant prescan spots the technologies and load numbers your brief names and flags the gaps a reviewable brief should not have. Derived from @sickn33/backend-architect (MIT).

Paste a rough bug report or feature idea and get a ready-to-file GitHub issue: a clean sentence-case title, the right issue type, a structured body, label suggestions, a readback you can correct and re-run, open questions and an honest confidence line. One click opens it prefilled in your own repo - no OAuth, no token, no repository access. The title linter, the type heuristic and the reply-to-reporter draft run free in your browser. Derived from @tldraw/write-issue.

Paste code — a module, a component, a handler, a query layer — and get a performance review: severity-ranked findings each with a corrected snippet, the quick wins worth shipping first, an optimized rewrite of your worst hotspot, and a five-area health check across data access, memory, CPU, network and rendering. A free instant prescan flags the classic anti-patterns before you run anything. Derived from @mastra-ai/performance-review.

Paste a service description and get a complete SLI/SLO package: user-centric SLIs with PromQL, justified SLO targets, an error-budget policy, multiwindow burn-rate alerts with verified arithmetic, gaps and open questions. Free error-budget calculator and burn-rate ladder run in your browser. Derived from the @wshobson/slo-implementation skill.

Paste a pull request - the unified diff, title and description - and get a presenter-ready walkthrough: segments ordered by understanding rather than by file, the exact hunk or function to show on screen for each beat, and the narration to speak over it in the author's voice. Generated files like lockfiles and API reports are filtered to a Not shown list. The diff scanner runs free in your browser. Derived from @tldraw/pr-walkthrough.

Describe the capability you want an AI agent to have - or paste a SKILL.md you already use - and get a complete, well-formed agent skill: trigger-complete frontmatter, an imperative body under the 500-line budget, a bundled-resource plan, and an honest review that raises open questions instead of inventing answers. The SKILL.md linter runs free in your browser. Derived from @tldraw/skill-creator.

Paste a resolved support ticket, a recurring question or a workaround write-up and get back a publish-ready knowledge base article: a searchable customer-language title, the right article type (How-to, Troubleshooting, FAQ, Known issue, Reference), category, tags and audience, a structured body built only from what the source establishes, publishing notes for the editor, and every unconfirmed fact routed to Gaps to fill instead of into the article. A publish-readiness verdict binds to those gaps: an unfilled gap rules out Publish as is. The source scan, title linter and article lint run free in your browser. Derived from @anthropics/knowledge-management (knowledge-work-plugins).

Paste a description of a system, a process, a data model or a set of interactions and get a real diagram file: downloadable as .drawio for draw.io and the hediet.vscode-drawio extension, as SVG, as Mermaid for a README, and as node and edge CSVs. Every diagram is drawn with one geometry - labelled boxes on tiers, joined by orthogonal connectors, with groups as dashed boundaries - so choosing flowchart, architecture, sequence, ER, UML class or network changes what the diagram contains, not the notation it is drawn in: there are no sequence lifelines, no ER crow's feet and no UML attribute compartments. The AI decides the shapes, what kind each one is, which layer it sits on, how they group, which connections exist and what each means; your browser does the deterministic half for free: layered layout, orthogonal routing, mxGraph XML, the SVG preview, and a structural check that catches duplicate ids, edges pointing at shapes that do not exist and shapes connected to nothing. A free browser-side scan reads the description first, scores it against the six diagram types, numbers every entity it finds and parses arrow syntax into a real graph - so a description that already states its arrows draws and exports itself at no cost. A derived work of @github/draw-io-diagram-generator.

Paste a GitHub Actions workflow and get a senior CI/CD engineer's review: a production-readiness posture, the workflow and job inventory, prioritized findings with corrected YAML fragments, quick wins and focus areas. A free instant prescan flags mutable action refs, missing permissions, script injection, secret leaks, missing timeouts and more. Derived from the @sickn33/github-workflow-automation skill.

Paste your research notes — code excerpts, design decisions, benchmarks — and get a publish-ready technical blog post draft: the angle (problem, insight, surprise), a full markdown draft, a seven-point quality checklist, and an honest no-story-yet verdict when the material is thin. Free readiness scan runs in your browser. Derived from the @tldraw/write-tbp skill.

Paste a task spec and the deliverable produced against it. Two independent AI reviewers check it against your rubric with no shared context - both must pass before it ships. A NICE verdict means it ships; NAUGHTY returns a deduplicated fix plan ordered blockers-first and tagged by which reviewer caught what, plus a reviewer-agreement metric, a paste-ready fix prompt, and a re-check that diffs the revision criterion by criterion. Derived from @affaan-m/santa-method.

Paste the code review feedback you received plus the code it is about, and get each comment triaged the way a senior engineer receives review: verified against the code, then routed to Implement (with the concrete change), Push back (with the technical reason), Clarify first (with the question to ask) or Verify in code (with exactly what to check) - plus a short, professional reply with zero performative agreement. Derived from @obra/receiving-code-review.

Paste a JavaScript or TypeScript module and get a complete Vitest or Jest unit test file: behavior-named tests, boundaries and failure paths covered, determinism hazards handled with fakes, plus a test plan and honest notes on what could not be tested. The surface scanner and hazard detector run free in your browser. Derived from @tldraw/write-unit-tests.

Paste UWP C# or XAML and get it back rewritten for WinUI 3 / Windows App SDK: every legacy API replaced with its modern equivalent, nine named findings marked MUST-FIX, CHECK or CLEAN, the manual steps that remain, compatibility notes, and a verdict with a confidence score. The legacy-API scan and a searchable UWP-to-WinUI 3 API map run free in the browser. Derived from @github/winui3-migration-guide.

Paste React, Vue, Angular, Svelte or plain HTML/JS source and get an i18n readiness audit scored out of 30: hardcoded strings, concatenation bugs, plural hacks, locale-unsafe formatting and RTL risks — plus a string-extraction plan with suggested keys and a downloadable en.json locale starter. The pattern scan runs free in your browser. Derived from @sickn33/i18n-localization.

Paste a documentation page and get a scored review: Readability, Voice, Completeness and Accuracy out of 10, a prioritized issue list quoting every defect with the concrete fix, and a rewrite of the worst passage. Paste the source code too and accuracy is verified against the truth. Free instant style lint runs in your browser. Derived from @tldraw/review-docs.

Paste Kotlin Compose UI code — a screen with its ViewModel, several files, or a grab-bag of composables — and get an expert review: a sound/refactor/rework verdict, severity-ranked findings with corrected Kotlin, a twelve-item Compose checklist, a five-area health check, and a full Refined.kt rewrite. A free instant prescan flags state-without-remember, exposed MutableStateFlow, GlobalScope work, unkeyed lazy lists, hardcoded colors and more before you run anything. Derived from the @affaan-m/compose-multiplatform-patterns skill (MIT).

Turn a command's --help or man output into a publish-grade tldr page. The free half is a real engine in the browser: a help-text parser that extracts every option and subcommand, a skeleton-page generator, a format lint measuring the tldr-pages rules (title shape, description block, More information line, 5-8 examples, the {{[-s|--long]}} placeholder conventions), a styled preview and .md export. The AI pass writes the examples people actually reach for, grounded in the pasted help text and re-checked by the same lint - every flag it uses that the help text never documented is printed by name. A derived work of @github/create-tldr-page (MIT, github/awesome-copilot).

Paste your git log, merged PR list or raw change notes and get a polished release-notes article: featured highlights, a migration guide for every breaking change, categorized changes with contributor credits, and the complete markdown ready to publish. An instant categorization scan runs free in your browser. Editorial method adapted from @tldraw/write-release-notes.

Paste your Terraform HCL and get an infrastructure-engineer review: a production-readiness posture, the block inventory, prioritized findings with corrected HCL fragments, quick wins and focus areas. A free instant prescan flags literal secrets, open ingress, public ACLs, wildcard IAM, missing version pins, local state, untyped variables and hardcoded AMIs while you type. Derived from @wshobson/terraform-module-library (MIT).

Paste Swift code and get a Swift 6.2 Approachable Concurrency migration review: data-race hazards, actor-isolation diagnosis, severity-tiered findings, a migrated rewrite in your own names, and a concrete build-settings migration plan. Free in-browser concurrency prescan before you sign in. Derived from @affaan-m/swift-concurrency-6-2.

Paste your Kubernetes YAML manifests and get a senior platform architect review: a production-readiness verdict, resource inventory, prioritized findings with concrete YAML fixes, quick wins, and the focus areas to tackle first. Free instant prescan catches missing limits, probes, :latest images, privileged pods and exposed services before you spend anything. Derived from @sickn33/kubernetes-architect (MIT).

Turn a sample API response into the response_semantics block and Adaptive Card templates that make Microsoft 365 Copilot render your plugin data as cards. The free half is a real engine: paste the JSON and it measures every field with type, sample and presence count, generates a grounded skeleton bundle, validates any bundle against the card rules AND against the sample — a binding that resolves on no item is named — and previews the exact card each item would produce. The AI pass shapes the template around your intent, and the same validator re-checks every binding it used, in public. A derived work of @github/mcp-create-adaptive-cards (github/awesome-copilot).

Paste GraphQL SDL and resolver code and get an expert API review: a sound/refactor/rework verdict, a five-area health check, severity-ranked findings with corrected SDL, a scored twelve-item checklist, and a full refined schema rewrite. A free instant prescan flags Int IDs, unpaginated lists, open introspection, missing depth limits and N+1 resolvers before you run anything. Derived from @sickn33/graphql (MIT).

Turn any web page into an importable Obsidian Web Clipper template. The free half is a real engine: paste the page's HTML and the browser extracts every meta tag, schema.org JSON-LD path and live CSS selector, generates a grounded skeleton template, validates any template against the Clipper import schema AND against the page, and previews the exact note it would clip. The AI pass shapes the template around your intent - and the same validator re-checks every selector, meta key and schema path it used, in public. A derived work of @sickn33/obsidian-clipper-template-creator (sickn33/antigravity-awesome-skills). Not affiliated with Obsidian.

Adjust every flexbox property on a live layout - direction, wrap, justify, align, gap on the container; grow, shrink, basis, align-self and order on individual items - add and remove items, and copy the generated CSS. For the properties where moving a control beats re-reading the spec. Fully client-side with no network calls. Imported Claude artifact. Source: https://github.com/simonw/tools (flexbox-playground.html, Apache-2.0).

Compare two JSON documents structurally - keys added, removed, changed - rather than as text, so reformatting and key reordering do not drown the real differences. Includes swap, clear and five worked examples. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (json-diff.html, Apache-2.0).

A working filter-chip prototype driven entirely from the keyboard: tab reaches every control, Enter and Escape behave correctly in the value popover, and focus returns to the chip that opened it instead of the top of the page. A reference for the focus-management details most filter-bar implementations miss. Fully client-side with no network calls. Imported Claude artifact. Source: https://github.com/simonw/tools (keyboard-filters.html, Apache-2.0).

Drop an animated GIF and extract every frame with its real metadata - delay, disposal method, dimensions, offsets - rendered one by one. For pulling a still out of an animation, or working out what timings a GIF actually contains when it plays wrong somewhere. Decoded in your browser with no third-party CDN code; nothing is uploaded. Imported Claude artifact. Source: https://github.com/simonw/tools (gif-dissector.html, Apache-2.0).

Paste formatted content from a document, web page, email or note and get clean Markdown back, with a live preview. Headings, lists, links, emphasis, code and tables survive; the styling noise does not. Incoming HTML is sanitised with DOMPurify before conversion. Runs fully client-side with no third-party CDN code. Imported Claude artifact. Source: https://github.com/simonw/tools (rich-text-to-markdown.html, Apache-2.0).

Three SVGs embedded as base64 data URIs, each shown rendered beside its decoded source - including one carrying a script element that renders fine and never executes, because an SVG loaded through an img tag is treated as an image with scripting disabled. A concrete look at why embedding untrusted SVG as an image differs from inlining it. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (svg-sandbox.html, Apache-2.0).

Press any keys and see what the browser actually reports: the key value (layout-dependent) alongside the physical code (not), for as many keys as you hold at once. The distinction is what decides whether a shortcut survives on someone else's keyboard layout. Modifier and navigation keys render as proper glyphs. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (keyboard-debug.html, Apache-2.0).

Paste formatted content from a web page and get every hyperlink back as text-plus-URL pairs, copyable as an HTML list, Markdown links, or plain text. A normal paste flattens links into bare words; this reads the clipboard's rich-text flavour and keeps the addresses. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (link-extractor.html, Apache-2.0).

Paste an epoch timestamp from a log, JWT or API response and read it as a full date - weekday, date, seconds, timezone - in both UTC and your local timezone, live as you type. Detects seconds versus milliseconds automatically from the digit count. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (unix-timestamp.html, Apache-2.0).

Paste a Pipfile.lock and get its pinned dependencies back as Pipfile entries and as requirements.txt lines, each with a copy button. Reads both default and develop sections, and skips entries with no pinned version rather than emitting broken pins. Handy for getting a requirements.txt out of a Pipenv project without installing Pipenv. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (pipfile.html, Apache-2.0).

Drop a JPEG and read its EXIF/TIFF orientation tag - the flag that decides whether a photo displays upright or sideways - with the raw value, its meaning, and debug detail like EXIF offset and file size. For settling which side of an upload pipeline is ignoring the tag. Parsed locally with DataView; the file is never uploaded. Imported Claude artifact. Source: https://github.com/simonw/tools (tiff-orientation.html, Apache-2.0).

Paste a bare millisecond duration from a log or benchmark and read it back as seconds, minutes, hours, days and weeks, plus a clock-style h:mm:ss.ms breakdown. Accepts numbers as you copy them - commas and underscores are stripped - and rejects non-numeric input rather than showing NaN. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (milliseconds.html, Apache-2.0).

Enter a file size (KB/MB/GB/TB) and a speed (MB/s, KB/s or Mbps) and see the transfer time as you type. Handles the bytes-versus-bits conversion that makes a 100 Mbps line move ~12.5 MB/s, not 100 - the usual source of a wildly optimistic estimate. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (transfer-time.html, Apache-2.0).

Convert text to and from HTML entities, live as you type. Escape mode handles the five characters that matter; unescape mode decodes named, numeric and hex references, which is what machine-generated HTML actually contains. Output is written with textContent, so your input is never rendered. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (escape-entities.html, Apache-2.0).

Paste terminal output from a Claude Code session and get clean prose back: the leading prompt character is stripped, terminal-wrapped lines are rejoined into paragraphs, whitespace runs collapse, and blank-line paragraph breaks survive. For quoting session output in an issue, commit message or doc without the wrapping artefacts. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (cleanup-claude-code-paste.html, Apache-2.0).

Paste text and see every smart quote, smart apostrophe, em dash and non-breaking space marked in place, with live counts per category. These are the characters that look fine in a document and then break a grep, a code block, or a diff. Covers the full U+2018-U+201F range, not just the common pair. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (curly-emdash.html, Apache-2.0).

A working demonstration of the BroadcastChannel API: open the page in several tabs, type in one, and the message appears in the rest instantly - no server, no backend. Same-origin and same-browser only, so nothing leaves your machine. Useful for seeing cross-tab state sync before building it yourself. Imported Claude artifact. Source: https://github.com/simonw/tools (broadcast-channel-chat.html, Apache-2.0).

Paste XML and get parse errors pinned to line numbers, highlighted in a numbered listing of your source. Checks well-formedness - unclosed tags, bad nesting, stray ampersands - not schema conformance. Uses the browser's own DOMParser, so the verdict matches your browser's. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (xml-validator.html, Apache-2.0).

Paste raw RTF markup - the {\rtf1\ansi... source itself - and get HTML with text colour, background colour and bold runs preserved as inline styles. The counterpart to a clipboard converter: for when you have RTF as text out of a file, an email source or a database column. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (rtf-to-html.html, Apache-2.0).

Choose a local folder and explore it in the browser: file tree, file and folder counts, total size, filter by type, a type-distribution chart, and click-to-preview for individual files. Built on <input webkitdirectory>; the browser hands over only what you select and nothing leaves the page. Imported Claude artifact. Source: https://github.com/simonw/tools (webkitdirectory.html, Apache-2.0).

A writing box that handles pastes properly: paste an actual file, or text over an adjustable threshold (1,000 characters by default), and it attaches below as a separate item instead of flooding your draft. Attachments show name, type and size, with inline previews for images. Files stay local; nothing is uploaded. Imported Claude artifact. Source: https://github.com/simonw/tools (pasted-file-editor.html, Apache-2.0).

Load a video, drag a crop box over the part you want to keep, and get the exact ffmpeg command - correct crop=w:h:x:y for the video's real pixel dimensions, plus an iPhone-compatible variant. The video loads locally as a blob for preview only; nothing is uploaded and no encoding happens in the browser. Imported Claude artifact. Source: https://github.com/simonw/tools (ffmpeg-crop.html, Apache-2.0).

Paste anything and see every MIME flavour the clipboard is really carrying - text/plain, text/html, text/rtf, image/png, vendor-specific types - with the raw payload behind each. Useful when a paste mysteriously loses formatting, or when writing paste-handling code. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (clipboard-viewer.html, Apache-2.0).

Paste a vision model's segmentation output - objects with box_2d bounds and base64 masks - and see the boxes and masks drawn over the image, with a switchable co-ordinate origin for when a mask lands mirrored or rotated. Ships an embedded example so it demos without an upload. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (mask-visualizer.html, Apache-2.0).

Pretty-print JSON that a normal formatter rejects - truncated mid-string, missing closing braces, or still streaming - with live updates as you type, a validity indicator and input/output character counts. Handy for half-arrived payloads and streamed structured output. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (incomplete-json-printer.html, Apache-2.0).

Paste a Swagger/OpenAPI JSON, tick the endpoints you want, and generate a trimmed spec containing just those paths plus the $ref definitions they depend on. Useful for cutting an oversized API spec down to something readable - or small enough to hand to a model. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (swagger-subset.html, Apache-2.0).

Build a JSON Schema visually: add typed properties and nested objects, give the schema a title and description, and copy the generated JSON. Handy for tool/function-calling definitions and structured-output contracts. State lives in the URL so the link restores your work. Runs fully client-side. Imported Claude artifact. Source: https://github.com/simonw/tools (json-schema-builder.html, Apache-2.0).

Draw pixel-perfect 24x24 icons with transparent backgrounds: click-drag to paint, erase, pick colours, or seed the grid from any emoji or character, then download a PNG. Your design is encoded in the URL, so the link restores your work. Imported Claude artifact, runs fully client-side. Source: https://github.com/simonw/tools (icon-editor.html, Apache-2.0).

Animated, side-by-side visualizer for eight classic sorting algorithms — bubble, insertion, selection, merge, quick, heap, radix and Timsort. Imported Claude artifact, runs fully client-side. Source: https://github.com/simonw/tools (sort-algorithms.html, Apache-2.0).