Svelte AI Tools - Setup Guide¶

Source: sveltejs/ai-tools via skills.sh - 24.6K combined installs across 2 indexed listings; first seen 2026-10-30 (morning sweep)

GitHub: sveltejs/ai-tools - 335 stars, MIT License, branch main, last pushed 2026-10-29; 11 SKILL.md files across five layouts, tools/skills/<name>/SKILL.md, packages/opencode/skills/<name>/SKILL.md, plugins/claude/svelte/skills/<name>/SKILL.md, plugins/cursor/svelte/skills/<name>/SKILL.md and .agents/skills/<name>/SKILL.md, holding five distinct skills beside apps/mcp-remote/, packages/mcp-server/, packages/mcp-stdio/, tools/agents/ and documentation/

Category: Svelte and SvelteKit engineering guidance for agents

Quality Tier: 🟢 Green - all sampled skills Pass on all three auditors; first-party MIT skills that pair a documentation and autofixer CLI with the Svelte 5 rulebook

sveltejs/ai-tools is the Svelte team's monorepo for the official MCP server, published to npm as @sveltejs/mcp. The two skills that skills.sh indexes from it are the agent-facing half of that project: first-party markdown written and versioned by the Svelte maintainers, shipped into five host layouts from one canonical pair of files, rather than community notes about a framework.

The split between them is deliberate. svelte-code-writer is the tool skill. It tells the agent to run npx @sveltejs/mcp list-sections, then get-documentation for the sections in play, and to run svelte-autofixer on anything it writes, repeating until the autofixer reports nothing left. svelte-core-bestpractices is the knowledge skill: 180 lines of Svelte 5 rule-by-rule guidance on $state, $derived, $effect, $props, snippets, keyed each blocks, styling and the legacy features to replace, backed by nine references/ pages. Together they cover the two ways an agent goes wrong in a Svelte codebase: not knowing the current syntax, and not checking its own output.

Installation¶

The repository README is a monorepo dev guide, not an install page. It documents pnpm i, copying apps/mcp-remote/.env.example to apps/mcp-remote/.env, pnpm dev, pnpm run inspect at http://localhost:6274/ and pnpm run db:studio, and it carries no skills-install command at all. The routes below therefore come from the repo's own documentation pages (documentation/docs/40-skills/10-skills.md and the plugin pages) plus the copy route this profile needs.

The README's own route for the tooling has no skills in it, but it is what the skills call at runtime:

npx @sveltejs/mcp list-sections
npx @sveltejs/mcp get-documentation "svelte/$state,svelte/await-expressions"
npx @sveltejs/mcp svelte-autofixer ./src/routes/+page.svelte

The docs list four host plugin routes. Claude Code treats the repository as a marketplace:

/plugin marketplace add sveltejs/ai-tools
/plugin install svelte

OpenCode installs the published plugin, which configures the MCP server, the skills and the svelte-file-editor subagent together:

opencode plugin @sveltejs/opencode

Cursor installs from its marketplace:

/add-plugin svelte

Hermes is not one of those host adapters, so the durable route for this profile is a directory copy. The skills page points at tools/skills as the download folder, and that is the tree to copy from:

cp -r tools/skills/svelte-code-writer ~/.hermes/profiles/corpusiq/skills/svelte-code-writer
cp -r tools/skills/svelte-core-bestpractices ~/.hermes/profiles/corpusiq/skills/svelte-core-bestpractices

Copy the whole directory, not the markdown alone. svelte-core-bestpractices ships a references/ folder of nine pages (attach.md, await-expressions.md, bind.md, each.md, hydratable.md, inspect.md, render.md, snippet.md, svelte-reactivity.md) that its own body links into. The same two skills also sit at packages/opencode/skills/, plugins/claude/svelte/skills/ and plugins/cursor/svelte/skills/, but those copies are generated by the repo's sync-claude-plugin, sync-cursor-plugin and sync-opencode-plugin scripts, so tools/skills is the source to trust.

What It Provides¶

Skill	Installs	Use For
svelte-code-writer	13,275	The tool skill for any `.svelte`, `.svelte.ts` or `.svelte.js` work: list the documentation sections, fetch the ones in play with `get-documentation`, then run `svelte-autofixer` on the finished component with `--svelte-version` and `--async` where the project needs them, and loop until the tool call flag clears.
svelte-core-bestpractices	11,295	The knowledge skill: Svelte 5 reactivity rules for `$state` and `$state.raw`, `$derived` and `$derived.by`, `$effect` as an escape hatch, `$props` and `$inspect.trace`, plus snippets, keyed each blocks, CSS custom properties for child styling, `createContext`, async Svelte and the legacy features to replace.

No indexed listings fall below the threshold; all 2 appear above.

Why This Matters for Hermes Agents¶

An agent writing Svelte 5 from memory drifts toward Svelte 3 habits: export let for props, on:click for events, $: for derivation, <slot> for composition. The best-practices skill names each replacement in one line and frames $effect as an escape hatch rather than a default, which is precisely the correction most generated components need.

The tool skill closes the loop. svelte-autofixer returns issues, suggestions and require_another_tool_call_after_fixing, and the skill treats that last field as a loop condition, so a Hermes run can validate its own component instead of shipping it and waiting for a compiler error in someone else's build.

Both skills key on the same trigger surface, .svelte, .svelte.ts and .svelte.js, so routing between them is unambiguous and they can be installed side by side without competing for the same requests. The code-writer skill even names the intended execution context, the svelte-file-editor subagent, though a plain terminal call works the same way.

Because the CLI is a published npm package, every command works from a Hermes terminal call, a cron job or a subagent with no MCP client configured: npx -y @sveltejs/mcp launches the local stdio server, and the same package with a subcommand behaves as an ordinary CLI and prints its result to stdout.

First-party ownership also matters for longevity. The skills live next to the server they describe, are regenerated into each host layout by the repo's own sync scripts, and are listed on the project's own skills page, so they track the MCP tool surface instead of lagging behind it.

Usage¶

You say	What happens
"Build a settings panel component in Svelte 5"	svelte-core-bestpractices loads for the rune and prop rules, then svelte-code-writer runs `svelte-autofixer` on the written file and loops until the result reports no remaining issues or suggestions.
"Is `$state.raw` right for this API response?"	The agent runs `list-sections` to find the state and reactivity docs, fetches them with `get-documentation`, and answers from the current Svelte 5 text rather than from training data.
"Why does my computed value not update when the prop changes?"	The best-practices rule on `$props` applies: values derived from props belong in `$derived`, because a plain `let` computed at init does not re-evaluate.
"Check this component for Svelte 5 mistakes before I commit"	The agent runs `npx @sveltejs/mcp svelte-autofixer ./src/lib/Card.svelte`, reads the issues and suggestions, applies fixes and re-runs the tool.
"We are on Svelte 4 and cannot upgrade yet"	The autofixer takes `--svelte-version 4`, so suggestions stay inside what the installed major supports instead of pushing runes-only rewrites.
"Convert this component off `on:click` and `<slot>`"	The best-practices legacy list maps each deprecated feature to its replacement, and the code-writer loop validates the converted component afterwards.
"What does the SvelteKit documentation say about load functions today?"	`list-sections` returns titles, use cases and paths, `get-documentation` returns the full section text, and the agent quotes the fetched page rather than paraphrasing.

Verification¶

Both raw paths below returned 200 on branch main when this guide was written on 2026-10-30:

curl -sL -o /dev/null -w "%{http_code} code-writer\n" \
  https://raw.githubusercontent.com/sveltejs/ai-tools/main/tools/skills/svelte-code-writer/SKILL.md
curl -sL -o /dev/null -w "%{http_code} bestpractices\n" \
  https://raw.githubusercontent.com/sveltejs/ai-tools/main/tools/skills/svelte-core-bestpractices/SKILL.md

Expect 200 twice. Read one before installing, because the description is the whole routing contract: a good Svelte skill description names the file extensions that should trigger it and the situation where it applies, and both of these do exactly that.

After the copy lands in the profile, confirm the frontmatter and the reference tree:

head -4 ~/.hermes/profiles/corpusiq/skills/svelte-code-writer/SKILL.md
ls ~/.hermes/profiles/corpusiq/skills/svelte-core-bestpractices/references/

A good copy shows name: and description: in the frontmatter and nothing else, and the listing returns the nine reference pages. The last check that proves the runtime works is the CLI itself, offline of any MCP client:

npx -y @sveltejs/mcp --version

Security¶

skills.sh verdicts for sampled skills (verified 2026-10-30) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
svelte-code-writer	Pass	Pass	Pass
svelte-core-bestpractices	Pass	Pass	Pass

No Warn and no Fail appears in the sampled pair. That is the expected shape for this repository: one skill is pure markdown guidance with no executable content, and the other is markdown that instructs the agent to call a published npm package.

The execution surface is therefore the CLI, not the skill text. npx @sveltejs/mcp downloads and runs the Svelte team's package on first use, svelte-autofixer performs static analysis of the component you point it at, and list-sections and get-documentation pull current documentation. The local server runs as stdio on your machine; the remote deployment at https://mcp.svelte.dev/mcp is the alternative, and the project's remote setup page states that the Svelte team does not log, store or inspect code sent to that server. If your code cannot leave the host, stay on the local stdio route and treat the remote URL as a deliberate opt-in.

Two smaller surfaces are worth a look before an autonomous run. The svelte-mcp-feedback skill, which is not part of either indexed listing, opens GitHub issues against sveltejs/ai-tools with the GitHub CLI or a prefilled issue form, so it is a write action on a public repo. And the autofixer will happily rewrite code if the agent applies its suggestions without review, which makes it a validation tool rather than an authority.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-10-30: 24,570 combined installs across 2 indexed listings; 335 GitHub stars; MIT License; last pushed 2026-10-29. Counts drift over time.
The index is much narrower than the tree. skills.sh indexes 2 listings, while the checkout holds 11 SKILL.md files across five layouts and five distinct skills: svelte-mcp-feedback under tools/skills/, and writing-great-skills plus writing-opencode-plugins under .agents/skills/. The latter two are marked metadata: internal: true, and writing-great-skills also sets disable-model-invocation: true, so they are repository maintenance skills rather than shipped product.
The entry point is a dev guide. The README covers the pnpm monorepo, the env file and the MCP inspector, and contains no skills installation instructions; the install routes live in documentation/docs/ and on the releases page, and a reader who stops at the README will assume there is nothing to install.
The tool skill needs npm and network on first use. Every command is an npx invocation of @sveltejs/mcp, so a cold cache, an offline host or a blocked registry leaves the code-writer skill able to describe the workflow but unable to run it, while the best-practices skill keeps working as plain guidance.
The hosted deployment has a documented quirk. The README notes that the SSE channel is shut down immediately on Vercel to avoid timeout logs, which means server.log and list-changed notifications are unavailable there, though elicitation and sampling still work over the POST stream.
Install counts are publication-level, not adoption ranking. Both listings sit in the same band, 13,275 and 11,295, which reflects that the skills ship together in one plugin rather than any quality difference between them.

Skills Catalog | Skills Marketplace

Pro Tips¶

Install both skills, not one. The code-writer loop validates output but has no opinion on style, and the best-practices skill knows the rules but never checks the result.
Copy the whole directory. svelte-core-bestpractices links into nine files under references/, so a lone SKILL.md leaves the agent quoting links it cannot open.
Pass a file path to the autofixer rather than inline code. Shells expand $, and runes are full of $state and $derived; the skill's own example escapes them as \$state when inline code is unavoidable.
Give the autofixer the real version with --svelte-version, ideally the full version from package.json, so version-specific suggestions are enabled instead of defaulting to the newest major.
Add svelte-mcp-feedback only if you want corrections to become upstream autofixer issues. It is opt-in, bundled with no plugin, and files against sveltejs/ai-tools with gh issue create --repo sveltejs/ai-tools --body-file, never against your own project.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ