Best of breed isn't dead. But capability rose so far this year that every product should ship every proven feature from day one. Proven features and mechanics are no-brainers now and cost roughly zero to execute, so there's no excuse to leave them out. Any high-growth SaaS can and should be replicated into opencore immediately, with no-hands ergonomics built in from the start: self-hostable core, optional managed layer, and a surface that agents can own through MCP and APIs so you barely touch the UI.
A year ago I posted a best-of-breed vendor stack. This is what actually held up.
The stack now
Split into themed tables so the picks are easy to scan.
AI and agents
| Category | Pick(s) | Why |
|---|---|---|
| Coding agents | Best token deal (all are viable) | Commoditised. They all work now, so optimise economics, not features. |
| Local dev agent manager | Conductor | Parallel isolated agents you review and merge. Always-on, multi-project. |
| Async tasks / bot manager | Cloud agent runners + your own MCP bot | Triggered from Linear or Slack, full MCP access. Still the least settled slot. |
| Personal assistant | Claude Code (/rc) | Quietly winning over OpenClaw and Hermes. It just manages its own things instead of running everything else. |
| LLM observability / evals | PostHog | Analytics, replay and LLM obs in one. Langfuse honourable mention. |
| AI gateway / model routing | LiteLLM, OpenRouter | LiteLLM is the default. OpenRouter is fine for bigger cases. |
| Sandboxes / code exec | Cloudflare Sandbox SDK | Isolated code execution on Containers, agent-callable over an API. |
| Web extraction | Firecrawl | Structured crawl and scrape agents can drive over MCP. |
Data and search
| Category | Pick(s) | Why |
|---|---|---|
| Data warehouse | Pub/Sub + BigQuery | Simplest reliable path to pull everything into one place. |
| ETL / CDC | PeerDB | Fast, Postgres-native CDC without the old pain. |
| Customer analytics | PostHog endpoints | Kill two birds with one stone. Tinybird honourable mention for bespoke. |
| Product analytics / flags / experiments | PostHog | One product across analytics, flags, replay and experiments. |
| Full-text / OOTB search | Typesense, Meilisearch | Opencore, self-hostable, drop-in. |
| Vector / graph / RAG search | pgvector + Neo4j / Apache AGE / Memgraph | Retrieval next to your data, no external search vendor. Qdrant if you outgrow pgvector. |
Platform and product
| Category | Pick(s) | Why |
|---|---|---|
| Auth (isolated) | BetterAuth | The most viable standalone option. OSS, TS-first, you own the surface. |
| Auth (full self-host) | Hexclave | Self-host the entire Clerk-like mechanic. More upfront work. |
| Payments | Stripe, Shopify | Kept up and still lead. Rare discipline at their scale. |
| Usage billing / metering | Stripe, Orb, Metronome | Usage-based pricing without hand-rolling meters. |
| Unosend | Own infra, generous free tier, better unit economics across many apps. | |
| Docs | docmd | Zero-config Markdown, native MCP, local models. Opencore, no SaaS tax. |
Infra and ops
| Category | Pick(s) | Why |
|---|---|---|
| Infra / hosting | Railway | Whole-canvas preview envs, PR-linked deploys, a surface agents can operate. |
| CI | GitHub Actions | Already integrated. Nothing else needs to be bolted on. |
| Secrets | Infisical | Self-hostable and agent-friendly. |
| Observability | Grafana + Prometheus | Still wins. Nothing agent-native beats it yet. |
| Background jobs / queues | Trigger.dev, Inngest | Durable async workflows agents can enqueue. |
| Agent work management | Linear + Slack | Trigger, close the loop, feed context back. |
| Chat | Slack (watch Buzz, Ano) | Incumbent holds. Open, agent-native challengers rising. |
The previous guard: still excellent, just not the current pick
None of these got worse. The model underneath them did. Most are still first-rate; they're just no longer the default in an agent-first, opencore world.
- Clerk (auth). Day-one gold in 2025 and still a strong product. By mid-2026 it reads as another SaaS surface: not agent-native, and the per-MAU tax bites once you scale. Great if you want managed and don't mind the dashboard.
- Cursor (coding agents). Set the bar for AI in the editor. Claude Code won the last year on the agentic loop and Conductor productised the fleet. The twist: post-acquisition the xAI stack and Grok are looking great, and I'm tempted to bet big here. Not decided yet, so it's on the watch list rather than the pick list.
- Mintlify (docs). First mover that made everyone's docs far better overnight, and still the quality leader. The catch is structural: it's the premium, pay-to-manage, closed version of what docmd now does in the open. When a late opencore entrant matches it on quality, open wins by default. Mintlify didn't regress; the model did.
- Resend (email). Best DX in the category. It forces paid tiers early once you run many apps and domains, which is where Unosend's economics take over.
- Notion (context/docs). Still lovely for humans. A peripheral context store is a chore now; what matters has to live in the repo or the tools agents already touch.
- Datadog (ops). So feature-dense it was hard to beat for getting started. Now partly displaced by folding observability into the PostHog and AI layer you already run, rather than by a better Datadog.
- Doppler (secrets). The nicer pure UX of the two. Infisical wins on self-hostability and agent-friendliness, which is what tips it in 2026.
The one clean exception: Statsig. This is an org story, not a model story. PostHog is simply crushing it, with a consolidated product and a strong captain at the helm, accelerating across analytics, flags, replay and AI obs at once. Statsig meanwhile got acquihired, its leadership taking over parts of OpenAI. Good for them. But that's basically that: the product isn't pushing the space anymore, and PostHog conquered the category around it.
Things to watch
Plenty is still in flux. The bets I haven't called yet:
- xAI and Grok, post-acquisition. The Cursor deal is looking great for the xAI stack. Tempted to bet big here, not decided yet.
- Personal assistant. Claude Code (/rc) is winning over OpenClaw and Hermes by just managing its own things rather than trying to run everything else.
- The async / bot-manager slot. The least settled of the lot: cloud agent runners plus your own MCP bot, triggered from Linear or Slack.
- Agent-native chat. Buzz and Ano against Slack.
- OOTB search. Typesense and Meilisearch consolidating the opencore lane.
The real shift
Two forces working together. Ergonomics that assume agents are the primary operators: MCP surfaces, always-on workspaces, context that lives next to code, zero dashboard tax. And opencore models that ship the high-traction ideas without locking you into pure SaaS rent.
The tools that survived became agent-first or made themselves self-hostable and MCP-ready. This is the stack that keeps moving forward.
As always, latest commentary at x.com/the_mewc.