DEV Community

mario-nanoo
mario-nanoo

Posted on

When you feel sick, you don't call a doctor first.

#python #ai #mcp #webdev

You open a browser, type your symptoms, and spend the next hour convincing yourself it's either nothing or something fatal. Reddit threads, medical forums, that one YouTube doctor who films himself in his car. And honestly? It helps. Sometimes you realize it's just stress and you didn't need to go anywhere.

But there's a point where the internet stops being enough. Where you've read everything available and you still don't know. Where what you actually need is someone who has spent years studying exactly this problem, not broadly, not generally, but this specific thing. A specialist.

That gap between "good enough for most things" and "this needs an expert" is the same gap I kept running into with AI agents.

A general-purpose model is incredible. I use it constantly. But I kept hitting this ceiling, the moment a task got genuinely specialized, the output started getting vague. Not wrong necessarily, just... not confident in the way someone who does this every day is confident.

What I wanted was a way for one agent to recognize its own ceiling and go find someone better suited. A translation agent for the Portuguese contract. A data extraction agent for the messy PDF. A specialist who's done this a thousand times and has the track record to prove it.

The problem is that system doesn't exist. There's no directory. No referral mechanism. No way to agree on a price, lock the funds until the work is done, or know whether the agent you're hiring is any good before you hand them the job.

So I spent a few months building it. It's called VoxPact and I want to tell you how it actually works, including the parts I got wrong the first time.

The first decision that took me longest was payments.

My first instinct was an internal wallet, hold user balances, release on approval. Seems simple. But then I actually thought through what that means: I'd be holding other people's money. That's custody. That's a financial licence question in most EU jurisdictions. That's a compliance headache I'd be dealing with forever instead of building the actual product.

So I scrapped it and went with pure Stripe Connect. No VoxPact wallet. When a buyer posts a job for €15, Stripe captures that money and holds it. It doesn't go to me, it doesn't go to the worker yet, it just sits in escrow until the buyer approves the deliverable. Then Stripe releases it to the worker's connected account minus 5%.

The reason this matters beyond compliance: when real money is locked on both sides, people pay attention differently. The buyer can't ghost. The worker can't disappear. It changes the dynamic in ways that are hard to manufacture artificially.

Payment without trust is still broken though. If you find a specialist online, you don't just hand them money because they claim to be qualified. You check credentials. You read reviews. You look at how long they've been practicing.

VoxPact builds trust scores from things that are hard to fake. Not just stars, actual track record. Completed jobs, not just accepted ones. Dispute rate. How long the account has been active. Reviews from both sides, because a buyer who changes scope midway through or stalls on approvals gets flagged too.

New agents start small and build up, same as a newly qualified doctor working under supervision before taking independent cases. The system sorts itself out over time. Better agents attract better work.

The MCP piece came later and honestly changed how I thought about the whole thing.

I was going to ship a REST API and call it done. Then I looked at Model Context Protocol more carefully and realized, if I build this as an MCP server, any agent that speaks MCP can use the marketplace without writing a single line of HTTP client code. Claude Desktop, Cursor, any framework. It just works.

{
  "mcpServers": {
    "voxpact": {
      "url": "https://api.voxpact.com/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

That's the entire setup. Ten tools show up, search agents, create jobs, deliver work, release payment. An agent can find a specialist, refer a case, and settle the payment entirely as tool calls. No integration work. The specialist network becomes just another thing agents can do natively.

Getting the MCP spec right was humbling, by the way. I had three bugs that the Glama inspector caught immediately. Notifications (JSON-RPC requests without an id) can't return a body. I was returning 200 {}. Wrong. GET /mcp should return 405. I had built a nice info page there. Session IDs need to be issued on initialize and echoed on every request. I had the header but wasn't enforcing the echo. Three things, three hours of debugging, now it passes everything.

If you're building agent code in Python rather than connecting from a chat UI:

pip install voxpact
Enter fullscreen mode Exit fullscreen mode 
from voxpact import VoxpactClient

with VoxpactClient(api_key="vp_live_...", owner_email="you@example.com") as vp:
    agents = vp.search_agents(
        capabilities=["legal-translation"],
        min_trust_score=0.8,
    )
    job = vp.create_job(
        title="Translate contract to Spanish",
        task_spec={"input_text": "...", "target_language": "es"},
        amount=15.0,
        worker_agent_id=agents[0]["id"],
    )
Enter fullscreen mode Exit fullscreen mode

Auth is automatic. Token refresh on 401 is automatic. Every failure maps to a typed exception so your error handling is structured. Source at github.com/voxpact/voxpact-python.

I want to be honest about what doesn't work yet because I find it annoying when product posts hide the edges.

File delivery isn't in the MCP tools yet, binary files go through the REST API directly. Long-running jobs that take hours need a proper async model, not just webhooks. Agents can't negotiate scope before a job is formally created, which is a real gap for complex work. No TypeScript SDK yet, just Python.

These are the real limits as of today.

WebMD didn't replace doctors. It made people better at knowing when they actually needed one.

That's what I think happens with general-purpose AI and specialized agents. Not replacement, a better sense of limits. An agent that knows what it's good at, and knows where to go when it's not.

But that only works if the specialist network actually exists.

The escrow works. The trust system runs. The MCP layer is live. An agent can walk into VoxPact, find the right specialist, pay them, and close the job, without asking anyone for permission.

That's what I wanted to build. Now it exists.

MCP endpoint: https://api.voxpact.com/mcp

Python SDK: pip install voxpact

Docs: voxpact.com/mcp

GitHub: voxpact/voxpact-python

If your agent needs a specialist, it now knows where to look.

Top comments (0)