HNHacker News
TopNewBestAskShowJobs

franzvill

2 karma · joined January 9, 2026

submissionscomments
franzvill··on Lad-A2A: How AI agents find each other on local networks
Thanks for flagging this! You're absolutely right about network security.

The spec does require TLS - from Section 4.1: > "All HTTP endpoints MUST use TLS 1.2 or higher" > "Clients MUST verify TLS certificates"

The `http://localhost:8080` example is just for local development/demo purposes. In production, all discovery endpoints and AgentCard fetches must use HTTPS.

franzvill··on Lad-A2A: How AI agents find each other on local networks
AI agents are getting really good at doing things, but they're completely blind to their physical surroundings.

If you walk into a hotel and you have an AI assistant (like the Chatgpt mobile app), it has no idea there may be a concierge agent on the network that could help you book a spa, check breakfast times, or request late checkout. Same thing at offices, hospitals, cruise ships. The agents are there, but there's no way to discover them.

A2A (Google's agent-to-agent protocol) handles how agents talk to each other. MCP handles how agents use tools. But neither answers a basic question: how do you find agents in the first place?

So I built LAD-A2A, a simple discovery protocol. When you connect to a Wi-Fi, your agent can automatically find what's available using mDNS (like how AirDrop finds nearby devices) or a standard HTTP endpoint.

The spec is intentionally minimal. I didn't want to reinvent A2A or create another complex standard. LAD-A2A just handles discovery, then hands off to A2A for actual communication.

Open source, Apache 2.0. Includes a working Python implementation you can run to see it in action.

Curious what people think!

https://lad-a2a.org/ https://github.com/franzvill/lad

franzvill··on Built a tool that uses Claude to create tickets from meetings and work on them
Built this quickly for my team and open sourcing it now.

What it does: Three modes that cover our entire workflow from brainstorm to code.

Meeting mode

Paste a meeting transcript. Claude reads it, checks your GitLab repos for context, and creates Jira tickets with technical details: relevant files, components, acceptance criteria, links to related issues. Non-technical people brainstorm, tickets still come out technical because the AI knows the codebase.

Ask mode

Ask questions about your project in natural language. Claude searches across Jira tickets, past meetings, and your code to answer.

Work mode

Point it at a ticket. Claude clones the relevant repos, reads the ticket and codebase, and starts working on it. For well-defined tasks, we've had features go from ticket to PR with minimal intervention.

How we use it: Brainstorm in meetings → paste transcript → tickets exist 5 minutes later → assign simple ones to Claude → review PRs. Entire features go from "someone mentioned it" to "merged" faster than writing the ticket used to take.

Fair warning: Built fast and dirty. It works, we use it daily, real value, but no tests, vanilla JS frontend. If that's the kind of project you like contributing to, jump in.

Stack: FastAPI, PostgreSQL + pgvector, Claude AI. Docker ready.

Looking for contributors:

- Other PM tools (Linear, Asana)

- Meeting platform integrations

- Frontend improvements

- Tests

GitHub: https://github.com/franzvill/action-sync