ZCode – Harness for GLM-5.2

Original Article ↗ Hacker News Discussion ↗

ZCode, a new desktop harness for the open-weight GLM‑5.2 coding model, is drawing interest as a Claude Code/Cursor-style alternative but with mixed reactions. Users report that GLM‑5.2 is capable and often less restricted than proprietary frontier models, yet slower and sometimes more expensive in practice, with opaque subscription quotas and frequent connection issues. A major fault line is trust: many developers prefer open-source, CLI-based or containerized harnesses and express strong concerns about installing a closed-source Chinese agent with broad system access, especially in professional or IP-sensitive environments.

Overall impressions of ZCode & GLM‑5.2

  • GLM‑5.2 is viewed as a strong coding model, sometimes comparable to high-end proprietary models for many tasks.
  • Several users note it is noticeably slower than Opus / GPT‑5.5 and that harness issues can cause long stalls on simple tool calls.
  • Some report ZCode as a “workhorse” they use heavily; others tried it and went back to OpenCode, which “felt smarter.”

UI, UX, and feature set

  • ZCode is GUI‑only today; multiple people want a CLI/TUI and compare it unfavorably to Claude Code’s TUI and OpenCode’s TUI/desktop.
  • The UI is repeatedly described as very similar to Codex, down to icons and layout.
  • It can connect to Docker and remote SSH hosts, which some see as essential for safe execution.

Cost, quotas, and peak pricing

  • Many are confused by opaque marketing like “base usage allowance included” and multipliers, though the app shows specific daily token quotas.
  • GLM‑5.2 is considered expensive and slower per task; some users burn through quotas quickly, especially on coding plans.
  • Peak/off‑peak coefficients (with peak defined as 14:00–18:00 UTC+8) discount GLM‑5.2 usage inside ZCode, roughly giving 1.5× effective allowance.

Trust, privacy, and geopolitics

  • A major thread questions installing a closed‑source Chinese desktop agent with full system access, citing Chinese national security laws.
  • Others argue US companies and intelligence pose similar or worse privacy risks; several conclude no major provider should be fully trusted.
  • Many recommend only running agents in containers/VMs, on separate users, or air‑gapped servers, regardless of country of origin.

Open source vs closed harnesses

  • Strong preference from many for open‑source harnesses (OpenCode, Pi, various TUI tools).
  • Debate over whether harnesses are truly “hard” or moat‑worthy; some say they are commodity wrappers, others claim they materially affect performance.

Alternatives and multi‑model setups

  • Multiple agnostic harnesses and routers (Pi, OpenCode, lmcli, Crush, role‑model, etc.) are mentioned as ways to mix GLM‑5.2 with other models.
  • Some use dual‑model or multi‑agent strategies (e.g., GLM for planning/security, faster models for implementation) to balance quality, speed, and cost.