My take on billion-context (ranxianglei/billion-context) after testing it with Claude Code:

• ⚠️ The launcher failed on the first try. bili claude returned a 403 from bili itself. Adding NO_PROXY=127.0.0.1,localhost fixed it on my box.
• ✅ After that it worked smoothly. Claude Code answered normally, and bili recorded the session (about 45,700 tokens of context).
• 🧪 I didn't test the big promise. The compression only matters in very long sessions, and I ran one short question.
• 🔐 It sits in the middle of your AI traffic as a local proxy with its own certificate. That's how it works, so only use it if you trust the code.
• 🧩 It's for power users. Lots of clients, ports and settings, and the docs are dense.

"It took one fix to get running, and it worked after that. Whether it really gives you endless sessions needs a much longer test."