2025年9月14日 の Show HN

Hey all! I recently gave a workshop talk at PyCon Greece 2025 about building production-ready agent systems.

To check the workshop, I put together a demo repo: (I will add the slides too soon in my blog: https://www.petrostechchronicles.com/) https://github.com/Aherontas/Pycon_Greece_2025_Presentation_...

The idea was to show how multiple AI agents can collaborate using FastAPI + Pydantic-AI, with protocols like MCP (Model Context Protocol) and A2A (Agent-to-Agent) for safe communication and orchestration.

Features:

- Multiple agents running in containers

- MCP servers (Brave search, GitHub, filesystem, etc.) as tools

- A2A communication between services

- Minimal UI for experimentation for Tech Trend - repo analysis

I built this repo because most agent frameworks look great in isolated demos, but fall apart when you try to glue agents together into a real application. My goal was to help people experiment with these patterns and move closer to real-world use cases.

It’s not production-grade, but would love feedback, criticism, or war stories from anyone who’s tried building actual multi-agent systems. Big questions:

Do you think agent-to-agent protocols like MCP/A2A will stick?

Or will the future be mostly single powerful LLMs with plugin stacks?

Thanks — excited to hear what the HN crowd thinks!

Hi HN, I’ve been working on https://datadef.io, a tool to help data team (engineer, architect, project manager) make sense of their data universe.

The problem:

- Data models (dbt, SQL, warehouses) often grow into a tangled mess of tables, joins, and undocumented assumptions. - Lineage is either scattered across tools or missing entirely. - Documentation is usually an afterthought (and gets outdated fast).

Datadef.io aims to fix that by providing:

- Interactive canvas to map tables, relationships, and indicators. - Automatic lineage visualization to trace dependencies. - Metadata management: define table/column-level details, ownership, and KPIs. - AI-generated documentation that stays in sync with your models. - Export/share features so asset managers, analysts, and other teams don’t get lost in spreadsheets or PDFs.

It’s still early, and I’d love feedback from the HN community. In particular:

What’s missing for you in lineage/metadata/documentation tools?

How would you want to integrate a tool like this into your workflow (dbt, Databricks, Power BI, etc.)?

I’d really appreciate your thoughts, feature requests, and criticism.

Thanks!

Demo: https://youtu.be/pnznDL9SZvI

PaperSync was a project was made by two CS freshmen, Matthew Li (me!) and Michael Li, in 24 hours during HackCMU.

At a high level, we built PaperSync to make reading research papers easier and more collaborative.

Users can reference any part of the paper, ask anything they want, and have other users reply, all within the paper itself!

If you are interested in our work, we would love to talk! Reach out to us at [email protected] or [email protected].

Several months ago, I posted in Show HN (https://news.ycombinator.com/item?id=42284412) about this project (previously named Privastead, now changed to Secluso). It's a privacy-preserving home security camera that uses OpenMLS for end-to-end encryption between the camera and the mobile app. The post received a good amount of attention and there were many good comments.

Since then, my project cofounder and I have made major improvements to the project. The project previously would act as a hub for an IP camera, which was otherwise closed source. But now, our camera software can also run directly on a Raspberry Pi (even one as weak as a Raspberry Pi Zero 2W), resulting in a security camera with a fully open source software stack. In addition, our Raspberry Pi-based camera can perform AI to detect people/pets/vehicles and send notifications. Moreover, our released camera binary can be verified using reproducible builds and our app now runs on iOS as well as Android. You can use this project to turn your Raspberry Pi into a fully functional and (more important) private security camera. Please check it out, use it, and provide us with feedback!

In addition, we built a prototype of a standalone home security camera using this open source project and a Raspberry Pi. Please check it out here (https://secluso.com/). It's not meant to replace the open source project, but to explore whether a plug-and-play camera could make it easier for people who are interested but don't have time to set up our project on a Raspberry Pi. We're curious if this kind of device would be useful to the community. If you'd like updates on our progress on that front, you can join our mailing list on the site.

Finally, we'd love to hear your feedback and ideas on how we can improve the project. And we always welcome contributions to our open source project.

Our site: https://secluso.com

vue-markdown-render is a tiny, client-side Markdown renderer for Vue 3. In browser-side benchmarks we see significantly lower CPU usage and latency vs Vercel’s Streamdown, with up to 100× speedups on large, code-heavy docs. Note: Streamdown targets server-side streaming; our results are for client rendering. Ideal for AI chatbot UIs.

Tired of your Go tests being slowed down by repetitive PostgreSQL database setup?

I built pgdbtemplate to solve exactly that. It uses PostgreSQL's native template databases to create fully migrated test databases in milliseconds instead of seconds.

The problem: Traditional test setup runs all your migrations for every single test. With a complex schema, this wastes minutes in every test run.

The solution: Create one "golden" template database. Every test gets its own isolated copy instantly via CREATE DATABASE ... TEMPLATE.

Benchmarks don't lie: - 1.5x faster for complex schemas (5+ tables) - 37% time saved creating 200 databases - 17% less memory usage - Performance gains increase with schema complexity

It's: - Thread-safe and built for t.Parallel() - Driver-agnostic: works with both pgx and pq - Integrates with testcontainers-go - MIT Licensed

I'd love feedback from the HN community, especially on: - The API design (ConnectionProvider/MigrationRunner interfaces) - Ideas for further performance optimizations - Security audit of the SQL formatting code

Check out the code and benchmarks on GitHub: https://github.com/andrei-polukhin/pgdbtemplate

Hi Hacker News, I’m part of the team behind

Kodosumi. We built it because deploying agentic services in production was much harder than building proof-of-concepts. Many frameworks let you build agents, but when you want them to run long tasks, scale across machines, or be observed/monitored reliably, things fall apart.

Kodosumi is our answer to that. It’s an open-source runtime that uses Ray + FastAPI + Litestar to let you:

Deploy “agents” and “flows” and expose APIs easily, with minimal config (just one YAML)

Scale horizontally for bursty workloads and long-running tasks without losing stability

Monitor realtime status, logs, dashboards, etc. so you can see what’s going on under the hood

Avoid vendor lock-in (you can plug in your own LLMs, vector stores, frameworks) and deploy on cloud, on-prem, Docker, Kubernetes, etc.

We’re still early, so some parts are in active development. But if you are working with AI agents — especially ones that need to run reliably over long durations / scale up under load — we think Kodosumi could save you a lot of infra pain.

Happy to answer questions, comparisons to e.g. LangChain / Ray Serve / custom setups, criticisms, etc.

Hey HN,

Bootstrapped founder here. I've got a bit of a story for you.

We started desplega.ai to build a sophisticated AI platform that could automate E2E testing. We spent the last few months talking to dozens of QA leaders, and trying to learn what are their actual challenges. We've got one consistent feedback from large teams: their daily reality is a living hell of slow, clunky tools.

We're talking about teams at major companies still managing tests on spreadsheets. Or they're stuck in a Jira instance so customized and slow you can “make a full pot of coffee” while a page loads (and that’s why they acquired arc! t3.gg said it first). On top of that, they're paying 2k/mo+!? for these tools that feel like they were designed in ‘05.

Soon, it became obvious that our AI tool was way too advanced for them, and why it was much easier for younger startups to start using us. But we didn't want to give up on them just yet so...

Because I grew up when the internet was still free, and I actually miss that a lot, we decided to create a free test management tool.

Our vision is still AI, but we learned AI is not the silver bullet large teams are wishing for. (We wrote something about it at https://harlequinetcie.substack.com/). Our hypothesis right now is that we can be that team building the right tools for each QA team, leveraging AI.

We would love to hear your thoughts on

(a) Should we make the project open-source? Any key features? (b) Would you ever trust an entity to do your QA first pass?

Cheers,

Hi HN,

In my spare time I built HumbleOp, an experiment in structured online debate. I’m not a full-time developer — this is a side project I’ve been working on evenings and weekends.

How it works:

A user posts an idea or opinion.

Everyone else can comment once, no threads.

The community votes. Importantly, a vote doesn’t mean agreement — it means you want that commenter to face the author.

When the timer ends, the top comment duels the author, one message at a time, publicly.

Others can react with likes or red flags. Too many flags and the winner is swapped with the runner-up.

Duels end when both click handshake, or if both top challengers are flagged out.

Direct duels are also possible, skipping the voting phase.

I built this because I was frustrated with how debates online usually go: endless replies, chaos, and low-effort takes dominating thoughtful ones. The strict rules are intentional: I wanted to see if structure could make discussions clearer and fairer.

Stack: React + Tailwind (frontend), Flask (backend), Postgres (database), deployed on Fly.io.

It’s live here: www.humbleop.com

It’s still a work in progress, so bugs are possible. Feedback is welcome — on the concept, the UX, or the tech.

Thanks for reading.

Warning: pre-release, unaudited, not for production use. (Though my password was generated with it)

Instead of saving secrets, it derives them on demand using domain + username + a short passphrase + a physical OpenPGP key (smartcard/YubiKey).

Passwords are reproducible but never persisted.

Currently tested only with RSA4096 on Windows + GnuPG 2.4.x.

Built on Browserbase with Stagehand

Hi HN, time for another round, I’m Lasse. For years I’ve been frustrated by classical ERP systems being closed, rigid and expensive to change (or just not changeable at all xd).

Last year I already posted my first attempt ("PORT", an OSS ERP), but due to better and better AI models, this approach already seemed outdated. So we iterared and today I’m sharing TNX API, an now open-source execution layer that lets AI read and write your business data, on your own server, with guardrails and logs.

We went from "full ERP" to an execution layer ("your database + AI" as a power employee). Everything runs on your own dedicated server, to support also legacy systems. Every AI action is logged and permissioned, admins and users have different rights. We leaned into tasks people actually do: generate PDFs (invoices/offers), bulk edits, charts, uploads -> extract -> write back and made them as simple as possible.

You simply ask in natural language -> AI crafts code -> our API Nexus checks/sanitizes -> Stargate executes that code against your DB -> result goes back to chat/forms/emailbot/shop. The system keeps audit logs for every step.

LINKS landing page: https://www.tnxapi.com/UI/LandingPage.php code: https://github.com/StultusEstQuiHocLegit/TNXAPI4

If you’re tired of ERPs that won’t bend to your business, kick the tires, file issues, and tell me what’s broken and I’ll fix it. I really hate how limited current ERP systems are and I’m going to do everything in my power to change that.

Always at your service, Lasse Tramann : )

PipelinePlus bundles common MediatR pipeline behaviors into a small, plug-and-play library to reduce boilerplate in .NET apps.

What’s included: - Validation (FluentValidation) - Caching (attribute-driven with opt-out) - Idempotency (deduplicate command handling) - Outbox (persist and dispatch events) - Performance (timing hooks) - Exception mapping (turn exceptions into consistent results/logs)

Motivation: In most MediatR projects, these cross-cutting concerns get reimplemented again and again. PipelinePlus aims to make them composable, testable, and easy to adopt without locking you in.

Install: dotnet add package PipelinePlus

Minimal usage: // Program.cs / composition root services.AddMediatR(typeof(Program)); services.AddPipelinePlus(); // registers the behaviors

  // Example request
  [Idempotent]
  public record CreateOrder(Guid Id, string Sku) : IRequest<Result>;

  // Example validator (FluentValidation)
  public class CreateOrderValidator : AbstractValidator<CreateOrder> {
    public CreateOrderValidator() {
      RuleFor(x => x.Sku).NotEmpty();
    }
  }

Notes: - Outbox/Idempotency plug into your own storage abstractions. - Works on .NET 8/9 and latest MediatR. - MIT licensed.

Repo: https://github.com/ilkanozbek/PipelinePlus Feedback welcome—API shape, naming, and ideas for the next behaviors (e.g., OpenTelemetry integration, Redis/distributed cache helpers).

I recently made a small open source project `EZLive`. It lets you host your private live video without a server with public IP address. It can also serve as a build block of a more complex media system.

- EZLive will create a localhost RTMP ingest server, allowing OBS to stream video. - Then EZLive converts the video into HLS files (.m3u8 + .ts). - HLS files are uploaded to S3-compatible storage (like MinIO, Wasabi, Cloudflare R2, etc.). - You can then watch the livestream with any HTML5 HLS player (like hls.js) in the browser.

Feature and drawbacks:

- Much lighter than Owncast — but no chat, no extra stuff. - Easy to set up — just one binary and a S3 bucket. - Scales easily — the S3 cloud storage handles the load.

If you just want a simple way to host your own livestream, this may work well for you.

This is a weekend side project I just finished. It’s still alpha, may have some bugs.

LookingForwardTo.com is a platform to discover, track and share things to look forward to. Movies, shows games, your own events and more.

I've been close to the concept of having things to look forward to ever since it basically got me out of a very dark time of my life back in 2017 (that and my late dog).

Last year I decided to try and combine that with some of my biggest passions and this platform is the result. I'm currently curating all content by myself, except for the release dates, which are automated. It's built with a Next.js frontend and Node.js + express.js + MongoDB with Mongoose backend. Feedback would be much appreciated!

A tool that checks if any website uses a paid Cloudflare subscription by examining their public /cdn-cgi/trace endpoint. Cloudflare only allows disabling Encrypted Client Hello (ECH) in paid plans, so sni=plaintext indicates a paid subscription. Works on any Cloudflare-proxied site.

Hello HN - solo dev and first timer, here.

Mirenku is a small desktop app to track what you’re watching. It’s local-first (runs offline, no telemetry) and meant to complement, not replace, MAL.

v0.3.2 highlights: quick add/+/– episodes, search/filter, JSON/CSV import/export, MAL OAuth2 (PKCE) + sync via protocol handler. Windows is primary; macOS/Linux builds are experimental.

Security notes: tokens are envelope-encrypted (Fernet AEAD DEK wrapped by the OS keyring); timestamped OAuth state; TLS 1.2+ only. SHA-256 checksums provided.

Links: - Download / landing: https://mirenku.app - Docs & security policy: https://mirenku.org - Source / license details: https://github.com/Aeturnis-Development-Labs-LLC/mirenku

Feedback I’m looking for: - Install friction on Win/mac/Linux - Import/Export edge cases - UI paper cuts (least clicks to update an episode?) - Concerns with the token storage approach

Happy to answer questions and fix rough edges quickly.