Ryan

For one slice of our backend we needed a fast, cheap chat model with strong Chinese-language output, and Z.ai's hosted GLM endpoint gave us sub-second latency without the cold-start or quota friction of self-hosted Hugging Face Inference Endpoints. Hugging Face is unbeatable for model discovery and fine-tuning experiments; Z.ai wins for "one endpoint, known SLA, done."

Slight twist on this one — Paean is our own platform, and Clide's agent actually runs through Paean to reach Claude. We use Claude (Opus and Sonnet) as the reasoning engine because its tool-use and long-context behavior set the bar for agentic workflows; Paean is the layer that gives each user their own memory, credits, MCP tool routing, and conversation history on top of it. So it's not Claude over Paean — it's Claude inside Paean, and that combination is what makes Clide feel less like a chatbot and more like a coworker.

Clide is a terminal app — so the dev loop that built it had to be terminal-native too. Claude Code runs right inside the shells I already had open, drives long-running tasks without stealing focus, and composes with any editor instead of locking me into one. Cursor is great if your center of gravity is the editor; mine is the terminal, and Claude Code matched that shape exactly. It also turned out to be the best dogfood possible — the more I used Claude Code to build Clide, the more obvious the design of Clide's own embedded agent became.