jcode claims 14x less memory than Claude Code

jcode publishes head-to-head memory numbers against Claude Code. For anyone running several coding agents on one machine, the number that matters is 117 MB, the baseline across ten concurrent sessions.

A single session with local embedding disabled runs at 27.8 MB PSS. The project claims that's 13.9 times less than Claude Code and 6 times less than jcode's own embedding-enabled mode.

Time-to-first-frame is the other headline figure. jcode logs 14.0ms against a competitor range of 383 to 3,436ms, a gap of roughly 27 to 245 times.

The harness is MIT-licensed and sits at 12,148 stars, with this specific comparison post adding 2,585 of those. It connects to Claude, OpenAI, Gemini, Copilot, Azure OpenAI, Ollama, and LM Studio through OpenAI-compatible endpoints. That makes it a drop-in front end rather than a model of its own.

The catch

These are vendor-published benchmarks, jcode comparing itself to its competitors with no independent verification. The methodology behind the 383-to-3,436ms competitor range isn't detailed enough to confirm it's an apples-to-apples test. Anyone running multiple AI coding sessions on one machine should treat these numbers as a lead worth reproducing, not a settled result.