pocketlantern

PocketLantern

Your AI agent sounds confident. It missed the blockers.

Prices shift, versions break, licenses change. PocketLantern is a blocker-aware decision layer for AI coding agents, grounded in source-linked, time-sensitive facts.

Try these queries

Ask your AI agent after installing PocketLantern:

"Which auth provider — Clerk, Auth0, or Cognito?"

⚠️ Cognito password hashes are permanently non-exportable
⚠️ Auth0 password export requires support ticket — Free tier excluded
⚠️ Auth0 Rules/Hooks EOL 2026-11-18 — Actions not portable
✅ Clerk has the most flexible migration path

"Should I upgrade to Next.js 16?"

⚠️ Sync API access fully removed — all dynamic calls must be awaited
⚠️ Custom webpack config breaks next build
⚠️ next lint removed — switch to Biome or ESLint CLI
⚠️ Requires Node.js 20.9.0+, TypeScript 5.1.0+, React 19.0+
✅ Plan phased migration — codemod doesn't cover webpack or middleware

"Can I use the OpenAI Realtime API?"

⚠️ gpt-4o-realtime-preview removed 2026-05-07
⚠️ Beta interface (realtime=v1 header) removed same date
⚠️ GA event schema incompatible — 4 event renames
✅ Start on GA from day one — or plan dual-stack migration

Three questions. Three blocker types. Same AI, different facts.

Quick start

Prerequisites: Node.js >= 22, an MCP client (Claude Code, Cursor, etc.)

npm install -g pocketlantern

Claude Code

In your project directory:

cd your-project
pocketlantern init

init registers the MCP server in ~/.claude.json and adds a one-line rule to your project's CLAUDE.md. Restart Claude Code (Ctrl+C, then run claude again), then ask your first query.

Cursor / Windsurf / other MCP clients

Add to your MCP config (Cursor: ~/.cursor/mcp.json):

{
  "mcpServers": {
    "pocketlantern": {
      "command": "pocketlantern-mcp"
    }
  }
}

Then add this line to your project rules (.cursorrules, Kiro steering file, etc.):

For technology decisions, upgrades, migrations, or licensing questions: before finalizing your answer, check the PocketLantern MCP server's search_cards tool for blockers in these bundled decision cards — your training data may be stale.

Reload MCP servers: Cursor — Cmd+Shift+P -> "MCP: Restart Servers". Windsurf — restart the editor.

More queries to try

4. "How did Vercel pricing change with Fluid Compute?" Billing split to Active CPU + Memory, shared-process concurrency breaks isolation

5. "Prisma or Drizzle for Edge/serverless?" Prisma v7 ESM + driver adapter breaking, Edge preview-only; Drizzle 0.30-1.0 also breaking

Verify installation

pocketlantern doctor               # check installation status
pocketlantern search "auth pricing" # search cards from CLI

How it works

PocketLantern is an MCP server. Your AI agent calls search_cards — and gets source-linked decision cards plus source-linked blocker warnings where available.

Agent asks question
  -> search_cards("clerk vs auth0 authentication")
  -> 3 cards + 5 blocker warnings (lock-in, EOL, export restrictions)
  -> Agent answers with source-linked constraints

Local cardset. Instant lookup. No vector database. No embeddings. Keyword search, card links, and source-linked blocker warnings where available. The value is in source-linked decision cards, not clever retrieval alone.

search_cards("auth", constraints:["serverless"]) -> cards + blockers
get_card("auth/clerk-vs-auth0-vs-cognito-2026")  -> full card with facts
get_related_cards(...)                           -> connected topics

What's included

Blocker-aware decision cards across 27 categories — plus source-linked blocker warnings from the bundled graph index. Ships with the npm package, works immediately, no network, no account:

Category	What it covers
auth	Vendor lock-in, migration pain, SSO, RBAC, passkeys
frontend	Next.js, React, Svelte, Vite, Angular upgrade blockers
database	Prisma, Drizzle, Postgres, Aurora, Neon, Supabase
ai	OpenAI API migrations, model pricing, batch vs streaming
serverless	Vercel, Cloudflare Workers, Lambda cost and runtime
infra	Hosting platforms, Terraform, Node.js LTS, Python EOL
backend	Supabase vs Firebase, Express, job queues, realtime sync
+ 18 more	deployment, devtools, testing, security, compliance, ...

What a card looks like

Each card is a structured YAML file with source-linked facts and official reference links:

id: auth/clerk-vs-auth0-vs-cognito-2026
title: Clerk vs Auth0 vs Cognito Under Current Pricing and Feature Changes
problem: Select an auth vendor given recent pricing shifts, MAU economics, ...
constraints: [cost-sensitive, low-ops, enterprise, compliance, serverless]
candidates:
  - name: Clerk
    summary: "Pro starts at $20/mo, includes 50,000 MRUs per app..."
    when_to_use: "Choose for small-team + low-ops + cost-sensitive SaaS..."
    tradeoffs: "Best DX and fastest implementation..."
    cautions: "Be precise about org-member limits..."
    links:
      - https://clerk.com/pricing
      - https://clerk.com/docs/guides/organizations/configure
  - name: Auth0
    # ...
  - name: Amazon Cognito
    # ...
tags: [auth, clerk, auth0, pricing, passkeys, b2b, compliance]
related_cards: [auth/sso-for-b2b-saas, auth/rbac-vs-abac-vs-rebac]
updated: 2026-03-14

See packages/schema/src/card.ts for the full schema definition.

Project structure

pocketlantern/
  packages/schema/       <- Card schema (zod + TypeScript types)
  apps/mcp-server/       <- MCP server (search, retrieval, tool handlers)
  apps/cli/              <- CLI (validate, search, init, doctor)
  packages/knowledge/    <- Decision cards + graph index (data package)
  docs/                  <- User guides & roadmap

Monorepo: pnpm workspace. Build order: schema -> knowledge (no build) -> mcp-server -> cli.

Development

pnpm build              # Build all packages
pnpm test               # Run tests (299 tests)
pnpm test:coverage      # Run with coverage report
pnpm lint               # ESLint
pnpm format:check       # Prettier check

Missing a topic? Request coverage in Card Requests.

Contributing

Contributions to the tool (schema, MCP server, CLI, search) are welcome. See CONTRIBUTING.md for guidelines.

License

MIT