Stack

What I actually use, what I dropped, and what I think about all of it.

Observability

langfuse.com

Using it for pipeline traces - it's crowded and heavy but it's the standard. Not using its evals or prompts features, just traces. I prefer custom-generated UI for evals. Built myself a CLI to pull traces so Claude Code can analyze them.

Changelog (1)
  • 2026-03-28Switched from Langfuse Cloud to self-hosted.
Using
lmnr.ai

Using it for browser agents. Feels better than Langfuse - has recording preview, SQL editor to pull traces, AI integration, and MCP support with Claude Code. More flexible and better thought-out for my use case, and fast loading too. Love it.

LLM Routing

github.com

My preferred routing SDK, but since that supply chain incident I've been slowly moving to native libs like google-genai for Gemini and so on.

github.com

Replaced by Bifrost

Heavy as fuck even on standby. Had many options I wanted to explore but never found the time. Switched to Bifrost and don't miss it.

Changelog (1)
  • 2026-05-03Replaced by Bifrost. LiteLLM proxy was using 5.5 GiB RAM with 8 Python workers.
Using
github.com
Changelog (1)
  • 2026-05-03Replaced LiteLLM proxy. Go binary, 111 MiB RAM vs 5.5 GiB.

CLI Tools

Using
llm.datasette.io

Very versatile to have around in the CLI. Use it for agents to call other LLMs fast. It logs stuff too.

github.com

My thinkboard. Obsidian canvas with Gemini connected via a custom plugin - I get a massive context window to branch ideas in different directions. Supports MCP, YouTube videos, images, image generation. Yes it costs money with Gemini, but prompt caching keeps it cheap - the API caches your conversation prefix so repeated turns only bill for the new tokens, not the full context every time. Turns a $2 brainstorm into a $0.30 one.

Using
warp.dev

I don't use the AI features - I just love how it handles organization. Tabs and panes (tabs within tabs) let me keep everything tidy without effort. The built-in file explorer is great, and the learning curve to start using panes and splits is basically zero. Was on Ghostty before this, and Windows Terminal when I was on Windows, but jumped back to Warp after they open-sourced it.

github.com

Open-source speech-to-text, works on Linux with OpenAI's Whisper model. No account needed, bring your own key. Polished UI for what it is. Two annoyances: doesn't pause system audio while recording, and Bluetooth headphones switch to hands-free profile (garbage quality) when the mic activates. Building a wrapper to handle audio profile switching automatically.

Automation

windmill.dev

I like it but I'm not great with managing workers, and the community version has some annoying limitations. May move to Kestra so people can see flows easily. It's way easier on the brain for me to just push to git and deploy.

Infrastructure

Using
coolify.io

Love it. Still use Cloudflare Pages for some stuff, but Coolify is versatile as hell.