Three days ago, Perplexity released Personal Computer. I think this changes what we mean by the word "computer."
Until recently, a computer was something you asked questions. Now it's something that does work. You wake up, and your always-on Mac mini has already spent the night ranking candidates, qualifying leads, or reconciling your books. Not a chatbot. A dedicated employee that lives on your desk.
Founders tinkering with OpenClaw, Claude Code, or Perplexity's new models keep arriving at the same conclusion. The software paradigm has shifted. For twenty years, we built software for humans. Now we have to build the infrastructure that agents run on.
Why SaaS is dead and MCP servers are next
The traditional SaaS model, dashboards, per-seat pricing, endless features, suddenly looks obsolete. The startups that win in 2026 will be the ones shipping MCP-native services. Clean APIs and custom skills that agents can discover, call with a high-level objective, and pay for only when they deliver a result.
Here's the simplest way I can explain it. Look at how classic SaaS compares to what's happening now.
Classic SaaS (2010–2025)
- User Humans clicking through a UI.
- Product UI-first features.
- Pricing Per seat or subscription tiers.
- Integrations Brittle, custom APIs.
- Governance User logins and access control.
- Moat Data lock-in and a well-designed interface.
- Distribution App stores and sales teams.
- Success metric ARR per seat and churn.
Agent-Native (2026)
- User Autonomous agents discovering tools via MCP.
- Product An MCP server, skill files, and orchestration.
- Pricing Per successful outcome.
- Integrations The universal MCP standard.
- Governance Scoped approvals, audit logs, kill switches.
- Moat Proprietary knowledge graphs and governance templates.
- Distribution MCP registries, ClawHub, agent marketplaces.
- Success metric Tasks completed and dollars saved.
If your product is just traditional SaaS with an AI chatbot bolted on, agents will ignore it. They might screen-scrape it as a last resort, but they'd rather have clean plumbing.
Get the full playbook
I wrote a detailed version of the framework below, including templates and starter files. Drop your email and I will send it over.
How to build an MCP server for AI agents
So how do you actually build for this? I've been running a loop: find gaps where agents fail, fill them. Eventually agents will run this loop themselves, but right now, human founders still do it better.
1. Pick a goal-seeking outcome. Find a repeatable task an agent could own end-to-end. Qualifying leads, screening engineers, things like that. If you can give Perplexity an objective and it gets 80 percent of the way there without you, that's a good candidate.
2. Map the MCP landscape. Look at registries, ClawHub, and GitHub. Where do agents get stuck? Usually they can pull raw data just fine but can't deeply evaluate it or enforce governance.
3. Design custom skills first. Don't build a full application yet. Write two or three SKILL.md folders, which are just natural-language playbooks with schemas and tool configurations. If you're building for hiring, one skill might evaluate GitHub repositories, another could draft personalized outreach.
4. Prototype on real hardware. Put your server on a dedicated Mac mini, which is becoming the standard agent runtime. Run fifty trials. Measure the success rate, cost per outcome, governance compliance. Most ideas die here. Good. It saves you time.
5. Price for agents. Agents don't want seats, so don't charge for them. Charge for outcomes. Thirty dollars per qualified shortlist, for example. You can add a premium for strict governance too, like unlimited audit logs and approval gates.
6. Make it discoverable. Publish the server to registries. Add webhooks so agents can trigger actions based on events. Write machine-readable documentation so agents can read your product the way humans read a website.
7. Build a moat. Multi-model routing helps balance reasoning and speed. But your real moat is proprietary data, like historical screening outcomes, and governance templates nobody else has.
8. Rank and ship. Score your opportunities by value, gap size, and readiness. Pick the best one and launch the MVP.
The gap in AI recruiting and technical hiring
I ran this process recently and found a big gap in technical talent screening.
During the Perplexity launch, the demo showed the computer prioritizing candidates and drafting emails. Agents are already good at sourcing resumes. But they're bad at deep technical evaluation, reading code signals, understanding domain nuance. And they lack ironclad governance.
This is an expensive problem. A bad engineering hire costs a company an enormous amount of money. Founders are already trying to use Personal Computer for recruiting, but they need something more reliable and more private.
SiftForge: an MCP server for technical hiring
So I started building SiftForge.
SiftForge is an MCP-native service that runs locally on your Mac mini. You give it an objective, say, finding a dozen senior React engineers in a specific salary band who are actively looking. It checks your local files and public signals, runs a deep technical evaluation, ranks candidates, and prepares the outreach. And it holds everything for human approval. Full audit logs. Remote kill switch.
It's local, discoverable by any agent, priced by outcome, with governance built in at the infrastructure level.
We're in the very early days of the Personal Computer era. The first wave of experimental agents is here. The second wave, polished, discoverable, outcome-driven services, is just getting started.
Join the SiftForge beta
First 200 founders get free beta on their Mac mini, plus a copy of the detailed playbook with templates and starter files.
We will email you the installer and your personal MCP endpoint.