contextfor.dev

One low-effort workspace for getting a product found: community listening, SEO, AI visibility (GEO) and content.

Vantage (the product behind contextfor.dev) replaces the separate listening, SEO, AI-visibility, link-building and content tools a founder would otherwise stitch together. You describe your product once, through a short AI interview, and the workspace keeps working in the background:

• GTM lane. Reads Reddit, Hacker News, X, YouTube, Product Hunt, Substack, GitHub, Stack Overflow, LinkedIn and your own feeds. It scores every conversation on fit, intent, evidence, momentum and timing, then emails up to five ranked conversations a day that are worth joining, each with a grounded draft reply. You write and post the reply. Vantage never posts for you.
• SEO and GEO lane. An AI agent over DataForSEO and Google Search Console data covers keywords, SERPs, backlinks and referring domains. It also checks whether answer engines (ChatGPT, Perplexity, Google AI Overviews) cite you. Paid calls need your approval before they run.
• Content lane. Turns what the listening and SEO lanes find into posts and social copy, grounded in your own knowledge base and run through a humaniser. Vantage writes the social content but doesn't schedule or post it. It points you to open-source schedulers instead.
• AI workspace. A single chat with three suggested prompts at all times. Answers come back as text plus visuals (rendered threads, backlink tables and graphs), and anything worth keeping is saved as an artifact.

How it works

1. Onboard from a URL. Paste your site and the interview drafts your monitoring profile (audience, problems, competitors, sources) for you to confirm.
2. Scan. An hourly Cloudflare Cron starts one durable Workflow per due workspace. Collectors fetch new conversations within a daily paid-provider budget.
3. Rank. Documents are embedded, deduplicated and scored. A learning loop re-weights ranking from your approvals, rejections and outcomes.
4. Ground and draft. Retrieval-augmented generation (RAG) pulls the most relevant passages from your knowledge base so a draft cites real evidence rather than inventing it.
5. Deliver. A daily digest arrives at the hour you choose, and the workspace holds the full queue, the SEO/GEO agent and saved artifacts.

Services and stack

Area	Service	Status
App	Next.js 16 (App Router), React 19, TypeScript, Tailwind v4	Live
Hosting	Cloudflare Workers via OpenNext: Cron, Workflows, Queues, R2, Browser Rendering, rate limiting	Live
Database	Neon Postgres + Drizzle ORM (additive migrations)	Live
Auth	Neon Auth (managed Better Auth), email OTP	Live
Billing	Polar: Pro at $19/mo (founding price), checkout, webhooks and customer portal	Live
AI	Vercel AI SDK 7 (tool calling with tool approval) through Vercel AI Gateway	Live
Embeddings + RAG	Workers AI embeddings (Qwen3) + Cloudflare Vectorize	Live
Embeddings + RAG (next)	Neon pgvector (halfvec) + pg_search hybrid search, Jina v5 embeddings and Jina reranker v3	Planned
Research	Tako Search via AI Gateway, with Linkup as fallback	Live
SEO / GEO data	DataForSEO (Labs, SERP, Backlinks, AI Optimization), cached per workspace	Live
Search Console	Google Search Console OAuth for first-party SEO data	Planned
Collection	Scavio, TinyFish, Firecrawl, HN Algolia, YouTube Data API, Product Hunt GraphQL, RSS	Live
Transactional email	Resend (digests and account email from updates.contextfor.dev)	Live
Newsletter	MailerLite for opted-in newsletters (news.contextfor.dev)	Planned
Bot protection	Cloudflare Turnstile on public endpoints	Live
Product analytics	PostHog: funnel events and browser error tracking, proxied first-party	In review
Worker errors	Cloudflare Workers Issues (no SDK)	In review
CI/CD	GitHub Actions (lint, typecheck, tests, migration drift, bundle-size gate, staging auto-deploy, gated production)	Live

Roadmap

Now (launch)

• Production acceptance of the unified workspace and the AI interview onboarding
• Reddit lane through Firecrawl search, with Scavio hydration for the daily top five
• PostHog funnel and Workers Issues alerts
• Signed one-click digest unsubscribe (RFC 8058)

Next

• Move RAG to Neon pgvector + pg_search with Jina v5 embeddings and the Jina reranker
• Google Search Console connection, plus a public SEO/GEO case study with ourselves as client zero
• Link enrichment and week-over-week topic trends
• MailerLite newsletter for opted-in users
• Landing-page lead capture

Later

• MCP server. Read your queue, opportunities and SEO data from Claude, Cursor or any MCP client, with a read-only tier on Free.
• Agent skills. A published SKILL.md so coding and research agents can use Vantage as a skill.
• Source discovery that suggests new communities, forums and newsletters for your niche
• More lanes (newsletters, podcasts, forums) and team workspaces
• Content calendar export to open-source schedulers (for example pendpost). Vantage still won't post for you.

Principles

• Never posts on your behalf. Every reply and social post is yours to send.
• Evidence over invention. Drafts are grounded in sources you can see. Fixture data is labelled as mocked, and production fails closed rather than showing it as a real lead.
• Spend is visible. Paid provider calls are budgeted, metered per workspace, and AI tool calls that cost money ask first.

Blog

Posts are Markdown files in content/blog/. Copy _template.md to <slug>.md, run pnpm blog:build, and commit both the post and the regenerated lib/blog/posts.generated.ts. Posts render to HTML at build time (no filesystem or Markdown parser in the Worker) and ship with canonical URLs, Open Graph tags, BlogPosting/BreadcrumbList/FAQPage JSON-LD, and entries in /sitemap.xml, /blog/rss.xml and /llms.txt.

Data model

Tables live in lib/db/schema.ts: workspace, source, document, lead, opportunity_outcome, plus the opportunity/budget/scan and distribution tables used by the sweep engine. Migrations in drizzle/ are additive only (expand/contract), so a Worker rollback always works against the live schema.

Collectors

Type	Order
rss	Direct fetch with conditional GET (ETag / If-Modified-Since) through fetchPublicText
substack	Publication feed via the RSS collector
hn	Algolia search + Firebase item enrichment
reddit	Scavio reddit.search → fixture
x	Scavio x.search → fixture
youtube	Scavio comments → TinyFish Fetch/Search → agent → YouTube Data API → fixture
producthunt	TinyFish Search+Fetch → homepage fetch → agent → PH GraphQL → fixture
github, stackoverflow, linkedin, alexandria	Native APIs and search

Every user-supplied URL goes through lib/http/public-fetch.ts: public http(s) only, no embedded credentials, no loopback/private/link-local/cloud-metadata hosts, manual redirect handling, timeout and streaming size caps — on top of Cloudflare's global_fetch_strictly_public DNS protection.

Multitenancy is enforced at the query layer. Coverage includes test/workspace-authorization.test.ts, test/queue-visibility.test.ts, and test/digest-actions.test.ts; refused digest mutations never reach the database.

Setup

pnpm install
cp .env.example .env.local
# Fill DATABASE_URL (pooled), NEON_AUTH_BASE_URL, NEON_AUTH_COOKIE_SECRET
pnpm db:migrate
pnpm dev

Optional: collector keys (SCAVIO_API_KEY, TINYFISH_API_KEY, PH_DEV_TOKEN, YOUTUBE_API_KEY), RESEND_API_KEY for digests, CRON_SECRET to protect the sweep, ALLOW_FIXTURES=true to permit sample data.

Set NEXT_PUBLIC_APP_URL to each deployment's own origin (including staging). Production-mode digest dispatch fails closed if it is unset; local development defaults to http://localhost:3000. Configure DIGEST_FROM_ADDRESS with a verified Resend sender and optionally DIGEST_FROM_NAME.

Digest dispatch isolates workspace failures: selection, delivery or last-sent recording errors count as failures without stopping other workspaces. Skipped or rejected sends never update the last-sent timestamp.

Digests select the active opportunity queue, not legacy leads: current monitoring profile, live same-workspace evidence, updated in the last 24 hours, ranked up to five entries. Email destinations use the same HTTP(S)-only policy as the app. The preferred UTC hour identifies a daily slot; the next cron catches up an elapsed slot across midnight, without sending it twice or less than 20 hours after the previous digest. First-time subscribers catch up the latest elapsed slot because preferences do not store an activation timestamp.

Each sweep handles at most 25 eligible workspaces, oldest attempt first, with a 90-second shared time budget and a 10-second email request timeout. hasMore reports batch/deadline deferral; later cron runs continue fairly, including when empty or failing workspaces remain due. A dedicated atomic lease serializes delivery per workspace; token-fenced completion cannot overwrite a newer claim.

Resend requests use a stable delivery idempotency key derived from the workspace and previous successful send. Resend retains keys for 24 hours. If email is accepted but recording fails, an unchanged retry is deduplicated within that window; a changed payload can be rejected as a conflict and remains failed. This is not an exactly-once guarantee across crashes or the provider retention window. Never manually replay uncertain deliveries without checking provider status.

Neon project

1. Create a Neon project and enable Managed Better Auth.
2. Copy the pooled DATABASE_URL and the Auth base URL into .env.local. Migrations use the direct (non-pooler) endpoint.
3. Generate a cookie secret: openssl rand -base64 32.

Scripts

Script	Description
pnpm dev / build / start	Next dev server / build / start
pnpm test	Vitest suite
pnpm typecheck	tsc --noEmit
pnpm lint	ESLint
pnpm db:generate / db:migrate / db:push	Drizzle migrations
pnpm cf:build / cf:preview / deploy	OpenNext build + Wrangler preview/deploy

Worker entry tests run separately: node --test test/worker-entry.test.mjs test/scan-run.test.mjs test/queue-consumer.test.mjs.

Deployment

Cloudflare Workers, configured in wrangler.jsonc: contextfor.dev and www.contextfor.dev custom domains, Workers static assets, nodejs_compat + global_fetch_strictly_public, observability, and the Browser / AI / Vectorize / R2 / Workflows / rate-limit bindings.

CI/CD lives in .github/workflows: CI runs lint, typecheck, tests, migration drift and build on every PR and push, plus a job that measures the Worker bundle against the 10 MB limit and fails the build if a Worker ever needs secrets again.

Before deploying this change, apply the additive digest migrations to the target database using its direct Neon endpoint. Automatic staging deployment does not run migrations; use the manual Deploy workflow with run migrations enabled for the intended environment. Set that environment's NEXT_PUBLIC_APP_URL before enabling digest delivery. Keep production deployment gated until the schema and runtime configuration are verified.

The Cloudflare Git integration runs independently of GitHub Actions. Verify its production branch, preview isolation and commands in the dashboard; the manual production gate in deploy.yml does not govern that separate integration.

pnpm deploy

Cron

Production schedules 0 * * * * (hourly, UTC) and staging 0 */3 * * * (every 3 hours). Digests go out at each workspace's chosen hour, and scans are gated by the plan's scan interval. worker-entry.mjs adds the scheduled handler missing from the adapter. With the configured SCAN binding, it starts per-workspace scan Workflows, requests cost rollup, then POSTs to /api/cron/digest. Without SCAN, it calls /api/cron/tick, which performs the scan, rollup and digest sweep itself. Both paths use WORKER_SELF_REFERENCE and Authorization: Bearer ${CRON_SECRET}; digest failures do not fail scan scheduling.

Worker script limits are 3 MB (Free) and 10 MB (Paid). CI measures .open-next/worker.js and fails the build above 10 MB.

Local development notes

• pnpm dev runs the Next dev server on Node, not Workers. Use pnpm cf:preview to exercise the real Worker runtime.
• On native Windows, opennextjs-cloudflare build calls fs.symlinkSync, which needs Developer Mode or elevation; without it the build fails with EPERM from the default linked pnpm layout recreating .pnpm virtual-store symlinks. This repo commits nodeLinker: hoisted in pnpm-workspace.yaml as the chosen fix: a hoisted install writes a flat node_modules with no symlinks, so OpenNext copies files instead of recreating symlinks. Alternatively, run the build on Linux (WSL, Docker, or CI). See docs/operations/2026-10-08-cloudflare-worker-build-and-domain.md.
• Operational caveat: a hoisted install copies every platform variant of native packages such as sharp's libvips instead of hardlinking them, so it uses more disk than the default linked layout. Budget for that in CI caches; it does not affect correctness of the build.
• The build requires no secrets. lib/auth/server.ts and app/api/auth/[...path]/route.ts both defer Neon Auth construction to request time, and CI has a job that fails if the build ever needs NEON_AUTH_* again.

Contributing

See AGENTS.md for the standard loop, the Superpowers workflow, available MCP servers (agent-browser, Supermemory, Greptile), Cloudflare and Neon operating rules, security invariants, migration conventions and how Greptile reviews PRs.

Ref orchestration

See .warp/REF_ORCHESTRATOR_SETUP.md for Warp Oz ↔ Ref Plans wiring.