mitul005

Nice work. How does this compare to running a Go-based MCP proxy like lazy-mcp? The core tradeoff I'm curious about is whether deferring tool schema loading at the protocol level is more reliable than filtering at the agent prompt level.

The container model is the interesting bit, AI can only touch what's inside $$...$$ blocks, the rest is frozen. But I'm not sure "programming language" is the right frame for it; it's closer to annotated scaffolding that delegates to an LLM. The real question is whether the caching via container hashing is reliable enough that you're not re-spending tokens on identical blocks across runs.

The iMessage auth problem is the right thing to be paranoid about. A channel that proves delivery doesn't prove identity someone texting through the same number or session could theoretically trigger approvals. Curious if you're thinking cryptographic verification or just out-of-band confirmation (like a PIN sent separately).

iMessage is the one that changes the category for me. "Agents that can text real people" is a different thing than a multi-agent demo. I've been hacking on something similar and the hardest problem wasn't orchestration it was figuring out when to bother a human vs. just deciding. How does auto-approval work, is it all-or-nothing or per task?

Cool will try

Looks interesting

The use case that immediately jumps out to me: being SSHed into a prod server during an incident and not wanting to context-switch to a browser just to eyeball a dashboard. Right now that context switch feels small but it breaks the flow at exactly the wrong moment. Curious how it handles Grafana's templating variables (the dropdowns that filter by host, environment, etc.) — those are load-bearing for most of our dashboards. If that's supported, this goes from "neat demo" to genuinely useful in incident response.

Good catch—C’s declarator syntax can definitely be misleading there, especially with multiple variables in one line.

add .claude folder and seeting.json where you can write all permission allowed not allowed.

Go’s io package works so well because interfaces are implicit and lightweight.

In C, you can mimic the shape with function pointers, but you lose a lot of the ergonomics and safety. Curious how far you can push composability before it becomes too cumbersome to use in practice.

Looks interesting

As an interviewer, the hard part isn’t whether someone uses assistance it’s whether they actually understand the solution.

Tools like this might push interviews toward more discussion-based or system design formats instead of live coding.

Check out my project: https://github.com/ferro-labs/ai-gateway

The traditional pathway for junior developers—education → entry-level job → mentorship—is weakening due to hiring freezes, rising expectations, and automation, leaving a gap with no coordinated solution.

ADRs are less about documentation and more about preventing “decision amnesia.”

Most systems don’t fail because of bad decisions, but because nobody remembers why those decisions were made.

Keeping them lightweight and versioned with the code seems key—otherwise they rot like any other doc.

Social systems need boundaries—context, norms, shared history. Feeds remove all of that and replace it with a single attention market.

Maybe the fix isn’t new features, but reintroducing locality as a first-class primitive.

Curious how it handles false positives though. Auto-generated fixes on a misdiagnosed alert sounds like a bad time. How much human review is expected in the loop before a fix gets applied?

Interesting, I want to try this for debugging our AI gateway when you're routing requests across 20+ LLM providers, the logs get noisy fast. Being able to filter by log level, jump between errors, and run SQL-style queries against structured logs in the terminal sounds like a huge time saver. No more grepping through multi-GB log files.