Armin Ronacher ⇌
@mitsuhiko
Followers
59K
Following
11K
Media
3K
Statuses
57K
Creator of Flask; A decade at @getsentry; Building new things — love API design & AI. Bypassing Permissions. Husband and father of 3 — “more nuanced in person”
Vienna, Austria
Joined February 2008
Recorded some updated thoughts on agentic coding tools if someone is curious.
15
89
754
And again. I’m not anti MCP. It just doesn’t work good enough yet.
7
0
27
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
879
3K
10K
Yesterday I showed Claude Code using lldb via tmux yesterday. Mario compared that to Claude using lldb via his own terminal MCP. Results are not entirely surprising.
2
0
10
Here another version of it debugging a logic bug with LLDB, this time in tmux and this time through a sub agent that it created itself.
Since @badlogicgames showed off his terminal remote control MCP I just wanted to show that Claude can use screen (and tmux) to remote control an LLDB run in a screen session. It doesn't do it perfectly, but it does a pretty decent job at it and I get value out of it.
2
0
29
Since @badlogicgames showed off his terminal remote control MCP I just wanted to show that Claude can use screen (and tmux) to remote control an LLDB run in a screen session. It doesn't do it perfectly, but it does a pretty decent job at it and I get value out of it.
6
4
56
I'm assuming this is some sort of unsolvable syntax trifecta: The desire to have JSON compatible literals in a syntax *and* braces for statements *and* everything being an expression.
4
0
11
Claude Code feels like a tool that is used by the developers for the purpose it's written. I'm sure most coding agents are used by their developers, but I wonder what the percentage of code changes to their agents come from using their own agent.
Both codex and cursor-agent (and Cursor IDE) have this flaw that they don't add an auto-timeout when they run bash scripts. Once they call a script that follows logs or watches for file changes and needs an explicit exit, it's over. Compare that to claude code which not only
8
5
86
My thoughts on GPT-5 so far are that I don't think architecturally, have bad ideas, bad taste, bad comms, and I'm not an opinionated engineer.
6
0
80
I want to record a video with me and another programmer who is not getting value out of AI. Goal is to have a good discussion about our approaches and learnings. Anyone interested?.
28
10
169
Lately, status in Europe is earned by trashing others (eg: the US). Europeans are stuck, soothing ourselves with the hope that the US will fail. That's toxic. The US may well be fine. We should use this opportunity showing that we can do better than them.
Europe in general and Denmark in particular suffers from the same sad disease. As much as I like the croissants here in the old world, I actually crave the unbridled optimism of America more when I've gone for too long without.
4
4
129
Saturday. 18:33. Nice weather. Vienna’s stores are closed. We are in a recession. You would think this country can no longer afford for force stores to close.
10
2
70
ChatGPT on Plus old: ~100 o3/week, ~100 o4-mini-high/day. Today: ~200 thinking/week. Not sure how this will play out, but has a o4-mini-high hardcore user on plus I'm guessing this will mean I will lose out compared to what I had :(.
3
2
26
gpt-oss is clearly trained on it and in vllm serve that tool would be provided. What is confusing to me though is that, how do I see all these search tool calls in opencode but not in codex with ollama?.
2
0
4