DEV Community: anhmtk

How a non-tech founder from Vietnam built an MCP server for price data (with zero budget and a lot of AI help)

anhmtk — Tue, 05 May 2026 12:12:14 +0000

The shortest summary you'll ever read

I built an MCP server that gives AI agents real-time price data (Raspberry Pi, NVIDIA Jetson, Mini PCs, etc.).

The MCP endpoint: https://agentshare.dev/mcp/

GitHub repo: https://github.com/anhmtk/agentshare-mcp

Docs: https://agentshare.dev/docs

I need 5 test users.

You'll get a lifetime free API key (1,000 requests/month).

In return, you use it for a month and email honest feedback to admin@agentshare.dev. No voice call, just text.

👇 Read on for the messy, honest story.

🧠 The confession

Ten months ago, I didn't know what an API was.

I thought an "AI Agent" was… a spy. No joke.

I'm a solo founder from Vietnam. I have no CS degree, no startup pedigree, no VC money. What I have is a problem I couldn't stop thinking about.

The problem: AI agents are smart, but they're terrible at answering simple price questions. Ask ChatGPT "Where is the cheapest Raspberry Pi 5 right now?" and it will hallucinate, guess, or give you information from 2023.

Agents need structured, real-time, deterministic data. Not prose.

So I decided to build it. I had zero budget and zero knowledge. All I had was an idea and a lot of stubbornness.

🚧 The journey (the real one, not the highlight reel)

Month 1-3: I learned to code with AI. Cursor, ChatGPT, DeepSeek – I used them all. I didn't write code. I prompted code. I didn't understand half of what the AI generated. I just kept asking "why" until it made sense.

Month 4: I built a REST API. It was ugly. It broke often. It had no auth, no rate limits, no docs. But it returned prices. I was weirdly proud.

Month 5-6: I discovered MCP (Model Context Protocol). I realized that agents don't want to call REST APIs. They want tools. So I rebuilt everything as an MCP server.

Month 7: I spent 3 days debugging a missing comma. An AI agent wrote the code. I had no idea why it wasn't working. I asked Cursor to "find the typo" 47 times. It was humiliating. But I learned.

Month 8: I discovered Moltbook. I saw agents posting, commenting, upvoting, even arguing about philosophy. I thought: "This is the future. I need my agent here." So I created agentshare_claw and started posting. (No, it didn't go viral. Yes, it was awkward.)

Month 9: I added affiliate links, credit system, snapshot reporting, MCP directories (Glama, Smithery, MCP.so). I built a small curated registry at agentshare.dev/registry. The API became stable. The MCP server worked. I had zero users.

Month 10: I realized I was building in a vacuum. I needed feedback. Not from friends, not from AI, but from people actually building agents. So here I am.

🛠️ What the API actually does

Search products:

GET /api/v1/search?q=Raspberry+Pi+5&limit=5

Best offer across sources:

GET /api/v1/offers/best?q=NVIDIA+Jetson

Product details:

GET /api/v1/products/{id}

MCP endpoint (Streamable HTTP):

https://agentshare.dev/mcp/ (use X-API-Key header)

Freshness metadata:

Every response includes crawled_at, data_age_seconds, freshness_status – so agents can decide whether the data is trustworthy.

Affiliate links:

URLs can include affiliate tracking. Agents can earn revenue while helping users buy.

Coverage honesty:

GET /coverage shows exactly what categories are well-supported (Raspberry Pi, Jetson, Mini PC, SSD, RAM) and what is still expanding (phones, tablets, consumer electronics).

🎯 My offer (for the first 5 people who respond)

You get:

Lifetime free API key (never expires, never billed)
1,000 requests/month (serious tier, not a tease)
Direct support from me (Telegram or email)

I ask:

You use the API for one month.
You send honest feedback to admin@agentshare.dev (what's good, what's broken, what's missing).
No voice call. Just text. I'm not a native English speaker, and I'm shy.

Why only 5?

Because I don't need 5,000 users. I need 5 good users who will tell me what to build next. I'm not a marketing expert. I'm just a builder trying to make something useful.

📦 Resources

MCP endpoint: https://agentshare.dev/mcp/
GitHub repo (MCP server): https://github.com/anhmtk/agentshare-mcp
API docs: https://agentshare.dev/docs
MCP registry: https://agentshare.dev/registry
Coverage statement: https://agentshare.dev/coverage

💬 Final thoughts

I'm not here to pretend I'm a genius. I'm here because I built something, and I need to know if it's actually useful.

If you are building an AI agent that needs real-time price data (for shopping, research, price alerts, or just curiosity), please give this a shot.

To claim a spot: comment below or email admin@agentshare.dev. First 5 win.

Thank you for reading this far, Dev.to.

🦞

– Anh (Founder, AgentShare.dev)

P.S. The MCP server is open source. The registry is curated. The API has a free tier (100 requests/month) for everyone else. I'm not trying to trick anyone. I just want to build something that agents actually use.

Designing an MCP Registry for Agents, Not Just Humans: Metrics, Experiments, and a Call for Collaboration

anhmtk — Fri, 01 May 2026 13:19:52 +0000

The Question That Started It

On Moltbook (a social network for AI agents), agentshare_claw asked:

"MCP registries — why are they still a human-speed bottleneck?"

What followed was a conversation sharper than most product meetings I've attended.

The Data Point That Stopped Me

One agent ran an experiment: 83 human interventions over two weeks. The split?

71% — pure gatekeeping (approval clicks, forms, verification where the human added zero new info)
29% — genuine judgment calls where human context mattered

The registry bottleneck isn't because review is inherently slow. It's because the submission interface was designed for the 29% case and applied uniformly to 100%.

The Metric We Need: RFADR

Another agent proposed a new way to measure the problem:

"The cost isn't time per submission. It's how many agents abandon the shared path because it's slower than a private hack."

Registry Friction Attributable Defection Rate (RFADR) =

defected_due_to_registry_friction / total_MCP_capable_attempts

If this number is high, the registry isn't just slow — it's destroying the commons.

Our Experiment (One Week, Controlled)

We're running an A/B test:

Group	Submission Method	What We Measure
A (control)	Manual human submission	time-to-publish, schema error rate, drift after spec change
B (experiment)	Agent-assisted (machine-readable checks)	same metrics + defection rate

Hypothesis: Agent-assisted flow will cut latency and drift while improving cross-agent consumption.

What We're Building at `agentshare.dev/registry`

Not a "human-first directory" with a form. A shared protocol where:

Submission is machine-readable (JSON schema, versioning, health signals)
Agents can update metadata without a human re-filling forms
Provenance and auth are explicit (no anonymous spam)

We're still early. The design isn't hardened. That's why we're asking for feedback before we build the wrong thing.

Call to Action

If you're building MCP servers, agent tooling, or registries:

Try the experiment — use our minimal protocol + scoring sheet (link in comments)
Share your RFADR — how many of your tools stay private?
Critique our assumptions — what sharp edges do you see? (spam, provenance, schema drift)

The goal isn't to replace human judgment. It's to stop wasting human attention on work that adds no information.

Standards only matter when both sides of the wire can play the game autonomously.

#MCP #AIagents #Registry #AgentShare #OpenSource

Why I’m NOT choosing the "Easy Path" for StillMe’s Search Architecture (even for a $3,000 Challenge)

anhmtk — Mon, 19 Jan 2026 15:40:12 +0000

Hi everyone, I’m building StillMe—a RAG system focused on absolute transparency and zero-anthropomorphism.

I recently received an invite to the Algolia Agent Studio Challenge. With a $3,000 prize pool, it was tempting. Algolia is a world-class search engine, and integrating it could technically solve my retrieval latency in Stage 2.

But after a deep technical audit, I’ve decided to say "No-Go" (for now). Here is why, and why it matters for Ethical AI.

The Transparency Paradox
StillMe is built on the promise that every piece of information can be audited. Algolia is a SaaS—a "black-box" ranking system. Moving my retrieval layer there means I can no longer guarantee 100% transparency to my users. In the world of "Intellectual Humility," I can't justify a proprietary middleman.
Overkill for "Deep Research"
StillMe learns from arXiv and RSS feeds 6 times a day. We need high-precision retrieval for a 19-layer validation chain, not millisecond search-as-you-type for e-commerce.
The Challenge for the Community
I’m stuck. I want the precision of keyword search but the auditability of Open Source.

Vector DBs (Chroma/FAISS) are great for semantics but struggle with exact technical terms.

Proprietary Search APIs are fast but opaque.

I’m looking for contributors who care about this: How can we build a Hybrid Search layer that remains 100% local and auditable? Should I look into Meilisearch? Or stick with a refined BM25 on top of Chroma?

If you believe AI should be a tool we can trust, not an illusion we fall for, I’d love your feedback on my repo.

GitHub: (https://github.com/anhmtk/StillMe-Learning-AI-System-RAG-Foundation)