Migrate docs from Mintlify to fumadocs

[masnwilliams]

Summary

Replaces Mintlify hosting with a self-hosted fumadocs (Next.js) app, deployable on Vercel.

What's preserved

• All 140 content pages unchanged — same file paths, same URLs (served at root, e.g. /browsers/standby). Content edits were limited to removing 3 dead snippet-import lines and adding frontmatter to one untracked testing doc.
• docs.json stays the navigation source of truth — lib/tree.ts converts its tabs/groups into the sidebar at build time, so the editing workflow doesn't change. Redirects ported to next.config.mjs.
• Mintlify MDX components still work — Info, Note, Tip, Warning, Card, CardGroup, Steps, Tabs, Accordion, Frame, Update, CodeGroup (incl. tab labels from fence titles, via a small remark plugin) are aliased to fumadocs equivalents in components/mdx.tsx.
• API Reference generated from the same Stainless OpenAPI spec URL via fumadocs-openapi (virtual pages, no generated files). Interactive playground is off (one flag in components/api-page.tsx to enable).
• GA4, Inter, brand colors, favicon, navbar links (Dashboard / Changelog / Careers / Sign up).

New (things Mintlify gated behind paid tiers)

• /llms.txt + /llms-full.txt (~970KB markdown dump)
• Raw markdown for agents: append .md/.mdx to any page URL
• AI page actions on every page (copy markdown, open in ChatGPT/Claude)
• Ask AI assistant (AI SDK + Anthropic) with docs search tool — needs ANTHROPIC_API_KEY set in Vercel; degrades gracefully without it
• Vercel preview deployments per PR

Known gaps / review notes

• API reference page URLs change (operation-id slugs like /api-reference/getProjects instead of Mintlify's generated paths) — worth eyeballing hardest in the preview
• Card icons (Font Awesome names) are dropped; search is local flexsearch/orama, not Mintlify's hosted search; API pages aren't in search yet
• Sidebar nesting >2 levels renders as collapsible folders, cosmetically slightly different from Mintlify
• Tests: bun run build compiles all 561 pages green, bun run lint green, key routes smoke-tested (content pages, API reference, llms routes, .md rewrites, redirects). No visual QA beyond that yet.

Deploy

Vercel: import repo, framework Next.js, bun install && bun run build, set ANTHROPIC_API_KEY, point docs.kernel.sh at it.

🤖 Generated with Claude Code

---

Note

High Risk
Full docs platform swap affects every URL, SEO, and deploy path; /api/chat exposes a paid LLM with abuse controls that must be configured correctly in production.

Overview
Replaces Mintlify with a fumadocs (Next.js) app (bun dev / bun run build, Biome lint). MDX stays under content/docs/; docs.json remains the nav source of truth via lib/tree.ts (sidebar, eyebrows, external links). Mintlify tags (Info, Steps, Tabs, etc.) are aliased in components/mdx.tsx so existing pages need minimal edits (snippet imports inlined, one testing doc gets frontmatter).

The app adds docs routes (catch-all MDX, separate OpenAPI pages from the same Stainless spec), site plumbing (sitemap, robots, OG images, GA4, BotID + firewall on /api/chat), and brand styling in app/global.css (lowercase UI, Kernel colors).

New product surface: Orama /api/search, Ask AI (floating panel + /api/chat with Flexsearch tool, Anthropic, origin/bot/rate limits, Braintrust), /llms.txt / /llms-full.txt, per-page .md agent routes, copy markdown and Open in ChatGPT/Claude actions. API docs use custom response dropdown UI and method badges in the sidebar; interactive playground is off by default.

Contributor docs (README, AGENTS.md) and .gitignore are updated for the Node/Next toolchain (bun.lock, .source, .next).

^{Reviewed by Cursor Bugbot for commit 4ccfdcc. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
Kernel	🟢 Ready	View Preview	Jul 21, 2026, 11:52 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs	Ready	Preview, Comment	Jul 25, 2026 11:04pm

[cursor]

Stale comment

Risk assessment: Medium

This changes the production delivery stack for the entire documentation site rather than only moving content: 195 files are affected, with new Next.js routing, OpenAPI rendering, search, raw-markdown endpoints, and deployment/build configuration. That gives the migration a broad user-facing blast radius.

The new POST /api/chat route also invokes a billable Anthropic model from client-supplied message history without authentication, rate limiting, or request-size/history validation. Although its tool access is limited to public docs, this creates material cost-abuse, prompt-integrity, and operational-load risk. The existing broken-link workflow still runs the Mintlify checker and is currently failing against the migrated tree, so migration validation isn't green.

I did not approve this PR because these cross-file runtime and operational changes meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This PR replaces the production delivery stack for the entire documentation site: 196 files are affected, with new Next.js routing, navigation generation, OpenAPI rendering, search, raw-markdown/OG endpoints, and build/deployment configuration. Although most documentation files are mechanical moves, the runtime and infrastructure changes have a broad user-facing blast radius.

The public POST /api/chat route also sends client-provided message histories to a billable Anthropic model without authentication, rate limiting, or request-size/history validation, which creates material cost-abuse and operational-load risk. The latest update adds vercel.json to select the Next.js framework; it doesn't reduce those existing risks and further confirms that this PR changes deployment behavior.

I didn't approve because the cross-file production migration and unauthenticated model-backed endpoint meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[vercel]

Additional Suggestion:

proxy.ts builds path-to-regexp templates as /{/*path} because docsRoute is "/", which only matches the literal root path, so header-based markdown negotiation never matches any real doc page.

[cursor]

Stale comment

Risk assessment: Medium

This PR still replaces the production delivery stack for the entire documentation site: 196 files are affected, including new Next.js page/API routes, navigation generation, OpenAPI rendering, search, markdown/OG endpoints, and Vercel/build configuration. Most content changes are mechanical moves, but the runtime and deployment changes have a broad user-facing blast radius.

The new public POST /api/chat route also accepts client-provided message histories and invokes a billable Anthropic model without authentication, rate limiting, or request-size/history validation. The latest update changes only the default model from Sonnet to Haiku; that may reduce per-request cost but doesn't remove the cost-abuse, prompt-integrity, or operational-load exposure.

I didn't approve because the cross-file production migration and unauthenticated model-backed endpoint remain Medium risk.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
Kernel	🟡 Building	–	Jul 21, 2026, 11:51 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[cursor]

Stale comment

Risk assessment: Medium

This PR replaces the production delivery stack for the entire documentation site: 196 files change, including new Next.js page and API routes, sitewide rendering/layout code, navigation generation, OpenAPI rendering, search, raw-markdown/OG endpoints, and Vercel/build configuration. Most MDX changes are mechanical moves, but the runtime and deployment cutover affects every docs page and has a broad user-facing blast radius.

The updated POST /api/chat now checks Origin and caps output at 2,048 tokens, which reduces cross-site browser abuse and per-request cost. It remains an unauthenticated, billable Anthropic proxy with no rate limiting or input/history size validation; direct clients can forge the matching Origin header and send high-volume or token-heavy requests. The latest layout and branding changes don't reduce the migration's sitewide deployment risk.

I didn't approve because the full-stack docs migration and remaining cost/operational exposure meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This PR replaces the production delivery stack for the entire documentation site: 196 files change, including new Next.js page and API routes, sitewide rendering/layout code, navigation generation, OpenAPI rendering, search, raw-markdown/OG endpoints, and Vercel/build configuration. Most MDX changes are mechanical moves, but the runtime and deployment cutover affects every docs page and has a broad user-facing blast radius.

The public POST /api/chat route remains an unauthenticated proxy to a billable Anthropic model. Its origin check and output cap reduce browser-based abuse and per-request cost, but there is still no rate limiting or input/history size validation, and non-browser clients can supply a matching Origin header. The latest commit adds only sidebar hover/active styling, so it doesn't reduce the migration or endpoint risks.

I didn't approve because the sitewide production migration and remaining cost/operational exposure meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

The actual diff changes 203 files and replaces the production delivery stack for the entire documentation site with a Next.js/Fumadocs application. It adds sitewide routing and rendering, generated OpenAPI pages, search, markdown/OG endpoints, a new dependency/build stack, and Vercel configuration. Most MDX changes are mechanical moves, but the runtime and deployment cutover affects every docs page and has a broad user-facing blast radius.

The PR also adds a public Anthropic-backed chat endpoint with a new system prompt, a server-side API key, BotID, dashboard-managed firewall controls, a 400,000-character request allowance, and Braintrust telemetry that logs user message content and client IP while instrumenting model output and usage. The latest update improves raw-body size enforcement and adds audit-log documentation, but it doesn't reduce the migration, privacy, cost, or operational surface.

I didn't approve because the sitewide runtime/infrastructure migration and the external model and telemetry integrations meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This PR replaces the production delivery stack for the entire documentation site: 204 files change (+4063/−449), with new Next.js/Fumadocs routing and layouts, OpenAPI rendering, search, LLM markdown/OG endpoints, build tooling, and Vercel deploy config. Most MDX moves are mechanical, but the runtime and hosting cutover affects every docs URL and has a broad user-facing blast radius.

POST /api/chat is a public, billable Anthropic proxy. Relative to earlier revisions it is better guarded (same-origin check, BotID deep analysis, Vercel Firewall checkRateLimit("chat"), ~400k-char body cap, 2,048-token output cap), but it remains unauthenticated and depends on dashboard-configured firewall rules outside this diff. Residual cost and operational abuse risk remains.

The latest merge-from-main and chat/index hardening do not reduce the overall migration risk below Medium: this is still a cross-file production-platform rewrite with new API surface and deploy behavior.

I did not approve because these infrastructure, sitewide UX, and model-backed endpoint changes meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This remains a sitewide production-platform migration, not a documentation-only change. The actual diff touches 206 files (+4,122 / −449) and replaces Mintlify delivery with a self-hosted Next.js/Fumadocs app: new App Router pages/layouts, next.config.mjs rewrites/redirects, proxy.ts, package.json/bun.lock, vercel.json, search, OpenAPI rendering, and agent/markdown/OG endpoints. Most MDX moves are mechanical, but the runtime and deployment cutover still affect every docs page.

The public POST /api/chat path still invokes a billable Anthropic model from client-supplied message history. Origin matching, BotID, Vercel Firewall rate limiting, and a large input-size cap reduce some abuse paths, but the endpoint stays unauthenticated and Origin checks are forgeable by non-browser clients. Braintrust instrumentation also logs chat input metadata (including client IP) when configured, adding an operational/privacy surface. The latest SEO commit (sitemap, robots, canonical/OG metadata) does not shrink that blast radius.

I did not approve because the full-site infrastructure rewrite plus the public model-backed chat route meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This remains a sitewide production-platform migration, not a documentation-only change. The actual diff touches 206 files (+4,139 / −449) and replaces Mintlify delivery with a self-hosted Next.js/Fumadocs app: App Router pages/layouts, next.config.mjs rewrites/redirects, proxy.ts, package.json/bun.lock, vercel.json, search, OpenAPI rendering, and agent/markdown/OG endpoints. Most MDX moves are mechanical, but the runtime and deployment cutover still affect every docs page.

The public POST /api/chat path still invokes a billable Anthropic model from client-supplied message history. Origin matching, BotID, Vercel Firewall rate limiting, and a large input-size cap reduce some abuse paths, but the endpoint stays unauthenticated and Origin checks are forgeable by non-browser clients. Braintrust instrumentation also logs chat input metadata (including client IP) when configured, adding an operational/privacy surface.

The latest updates add production SEO surface (sitemap, robots, canonical/OG metadata) and spread shared Open Graph fields into per-page metadata. That does not shrink the migration blast radius and expands indexing/social-preview behavior for the cutover.

I did not approve because these infrastructure, sitewide UX, model-backed API, and SEO cutover changes meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Stale comment

Risk assessment: Medium

This remains a sitewide production-platform migration, not a documentation-only change. The actual diff touches 206 files (+4,161 / −449) and replaces Mintlify delivery with a self-hosted Next.js/Fumadocs app: App Router pages/layouts, next.config.mjs rewrites/redirects, proxy.ts, package.json/bun.lock, vercel.json, search, OpenAPI rendering, and agent/markdown/OG endpoints. Most MDX moves are mechanical, but the runtime and deployment cutover still affect every docs page.

The public POST /api/chat path still invokes a billable Anthropic model from client-supplied message history, with a new system prompt and server-side API key. Origin matching, BotID, Vercel Firewall rate limiting, and a large input-size cap reduce some abuse paths, but the endpoint stays unauthenticated and Origin checks are forgeable by non-browser clients. Braintrust instrumentation also logs chat input metadata (including client IP) when configured, adding an operational/privacy surface.

The latest update only wires per-page OG image generation for API-reference routes (app/og, getApiPageImage). That is a narrow SEO/preview correctness fix and does not shrink the migration blast radius.

I did not approve because these infrastructure, sitewide UX, model-backed API, and SEO cutover changes meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Risk assessment: Medium

This remains a sitewide production-platform migration, not a documentation-only change. The actual diff touches 206 files (+4,178 / −449) and replaces Mintlify delivery with a self-hosted Next.js/Fumadocs app: App Router pages/layouts, next.config.mjs rewrites/redirects, proxy.ts, package.json/bun.lock, vercel.json, search, OpenAPI rendering, SEO (sitemap/robots/canonical/OG), and agent/markdown endpoints. Most MDX moves are mechanical, but the runtime and deployment cutover still affect every docs page.

The public POST /api/chat path still invokes a billable Anthropic model from client-supplied message history, with a new system prompt and server-side API key. Origin matching, BotID, Vercel Firewall rate limiting, and a large input-size cap reduce some abuse paths, but the endpoint stays unauthenticated and Origin checks are forgeable by non-browser clients. Braintrust instrumentation also logs chat input metadata (including client IP) when configured, adding an operational/privacy surface.

The latest update only hardens chat JSON parsing (400 on malformed bodies) and skips external-link stubs in llms endpoints. That does not shrink the migration, cost, or operational blast radius.

I did not approve because the full-site infrastructure rewrite plus the unauthenticated model-backed endpoint meet the Medium-risk threshold.

  

_{Sent by Cursor Automation: Assign PR reviewers}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

There are 6 total unresolved issues (including 5 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4ccfdcc. Configure here.}

[cursor]

Ask AI draft never restores

Medium Severity

The Ask AI input reads localStorage inside the useState initializer with a server guard, but on SSR that initializer runs with window undefined and the state is fixed to ''. React reuses that state on hydration and does not re-run the initializer on the client, so a saved draft written via onChange never appears again after refresh or navigation.

 

^{Reviewed by Cursor Bugbot for commit 4ccfdcc. Configure here.}