Langfuse
Open-source tracing, evals, and prompt management for LLM apps
About Langfuse
Langfuse is the observability layer most LLM apps are missing. It traces every call your app makes — prompts, completions, tool calls, latency, token cost — and gives you evals, datasets, and prompt management on top, so you can answer "why did it do that?" without re-running the whole thing by hand.
What you get
- Tracing — nested spans across chains, agents, and tool calls, with cost and latency attached to each step.
- Evals — run model-based or custom scorers over datasets so prompt changes get measured instead of vibed.
- Prompt management — version prompts outside your codebase and roll them forward without a deploy.
- Integrations — OpenTelemetry, LangChain, LlamaIndex, LiteLLM, and the OpenAI SDK, so instrumenting is usually a few lines.
Good for
Anyone running an LLM feature in production, or anyone whose API bill is growing faster than their understanding of it. Self-host the whole thing or use the managed cloud; the core is MIT-adjacent open source and the project came out of YC W23.
Repo: github.com/langfuse/langfuse.
More Github
An agent skill that stops your AI generating generic, obviously-AI interfaces
taste-skill
AI coding agents converge hard on the same visual defaults: the same gradients, the same card layouts, the same spacing. taste-skill is a skill you load into your agent to push it away from those defaults and toward design decisions that look deliberate. Worth trying if everything you ship looks like it came out of the same template, because to a large extent it did.
A working coding agent built from zero in about 600 lines of TypeScript
pi-from-scratch
A minimal agent harness written from scratch, small enough to read in one sitting. No framework, no abstraction layers, just the loop: prompt, tool call, result, repeat. The fastest way to stop treating your coding agent as magic. Once you have seen the whole loop in 600 lines, the behaviour of the big harnesses stops being mysterious and starts being debuggable.
Node-based, tmux-backed terminal manager for running coding agents in parallel
nodeterm
Running three or four coding agents at once quickly becomes an exercise in losing track of which terminal is doing what. nodeterm gives each agent session a node in a visual graph, backed by tmux, so parallel work stays legible. Useful the moment you stop running one agent at a time and start treating them as a small team you supervise.
Turns a failed agent trace into a regression test that blocks the PR
Tracely
Most agent eval suites test the failures you imagined. Tracely tests the ones that actually happened. It is trace-native CI/CD for AI agents: when an agent fails in production, that trace is captured and converted into a regression test that blocks the pull request next time. That inversion matters because the failure modes worth guarding against are rarely the ones you predicted at design time. If your agent already embarrassed you in front of a user, this is the tool that stops it happening twice.
book-to-skill
book-to-skill converts a technical book PDF into a ready to load Claude Code skill, so the reference you never finished reading becomes something your agent can consult while it works. Point it at a PDF and get back a structured skill directory with the material chunked for retrieval. A neat way to turn a shelf of unread technical books into working context.