By Z.ai just dropped ZCode — a free desktop AI coding environment powered by GLM-5.2, an open-weights model that lands within one percentage point of Claude Sonnet 5 on coding benchmarks. It undercuts Cursor and Claude Code on price, and the app is available now on macOS, Windows, and Linux. There is a catch — and it is not a small one.
Not an IDE. An ADE.
ZCode does not wrap an editor with an AI panel bolted to the side. Z.ai calls it an Agentic Development Environment — the agent conversation sits at the center, and everything else (file manager, integrated terminal, Git panel, live browser preview) is arranged around it. The distinction changes how you work.
The primary interface is the /goal command. You type an objective — /goal refactor authentication module to use JWT — and ZCode plans the work, edits files, runs terminal commands, reviews output, and iterates until it finishes or hits a decision point. There is no “press Tab to accept a suggestion” loop. You define outcomes, not individual edits.
Five autonomy levels let you control how much ZCode does on its own:
- Plan Mode — shows the full plan before touching anything
- Confirm Before Changes — approval required for each file edit
- Auto Edit — files change automatically, commands still need approval
- Default Mode — balanced prompts, moderate autonomy
- Full Access — minimal interruptions, maximum throughput
Custom subagents are defined as Markdown files in ~/.zcode/agents/ — a pattern familiar to anyone using Claude Code. ZCode also imports MCP server configurations directly from existing Claude Code or Cursor setups, so your tool integrations carry over without reconfiguration.
How Good Is GLM-5.2 Actually?
GLM-5.2 is Z.ai’s flagship open-weights model, released June 13, 2026 under an MIT license. It is a mixture-of-experts architecture with 744 billion parameters, 40 billion active per inference pass, and a 1-million-token context window. The model weights are publicly available and inspectable — a rare property in a model performing at this level.
The benchmark numbers are competitive:
| Benchmark | GLM-5.2 | Claude Sonnet 5 | Claude Opus 4.8 |
|---|---|---|---|
| SWE-bench Pro | 62.1% | 63.2% | 69.2% |
| FrontierSWE | 74.4% | — | 75.1% |
| Terminal-Bench 2.1 | 81.0 | — | — |
GLM-5.2 is the first open-weights model to cross 80% on Terminal-Bench 2.1, a benchmark designed for agentic terminal workflows — exactly what ZCode runs on. On FrontierSWE, which tests long-horizon coding tasks where context retention matters most, GLM-5.2 is 0.7 points behind Opus 4.8. That gap costs significantly less: GLM-5.2 runs at $1.40 per million input tokens versus Sonnet 5’s introductory $2 and Opus 4.8’s standard rate.
What ZCode Costs
The app itself is free. Agentic work powered by GLM-5.2 requires a subscription after a five-day trial that provides 5 million tokens per day (3M GLM-5.2 plus 2M GLM-5-turbo).
| Plan | ZCode (GLM Coding) | Claude Code | Cursor |
|---|---|---|---|
| Entry | Lite: $16.20/month | from $20/month | Pro: $20/month |
| Mid | Pro: $64.80/month | Max: $100/month | Business: $40/month |
| High | Max: $144/month | Max: $200/month | Enterprise |
ZCode’s entry tier is the cheapest among the major agentic coding tools at $16.20 per month. The Pro tier at $64.80 undercuts Claude Code Max by more than a third. The differentiator at every tier is GLM-5.2 access — a model that benchmarks at roughly 95% of Sonnet 5’s SWE-bench score at 70% of the per-token cost.
The Data Trade-Off — Read This Before Installing
ZCode routes everything through Z.ai’s infrastructure, which operates under Chinese jurisdiction. China’s National Intelligence Law of 2017, Article 7, requires Chinese organizations to cooperate with state intelligence work on demand. That obligation applies to Z.ai regardless of where its servers are physically located, regardless of any enterprise privacy agreement, and regardless of whether a specific request has been made.
ZCode does support bring-your-own-key (BYOK) access for Anthropic, OpenAI, DeepSeek, OpenRouter, and others. This matters less than it sounds. BYOK changes which model generates the text completion. It does not change which server sees your code. Your codebase, file structure, terminal output, and Git history are still processed by ZCode’s orchestration layer running on Z.ai’s servers. The model inference endpoint changes; the code routing does not.
Self-hosting GLM-5.2 closes this gap but opens a different one. Running the full 744B parameter model requires approximately 1.5TB of GPU memory — roughly 19 H100 80GB cards. That hardware investment is not realistic for most engineering teams, and the cloud GPU cost to run it continuously is not clearly cheaper than a subscription.
One additional operational concern flagged in community discussion: ZCode agents do not retain context between sessions. On a large codebase, the time spent re-explaining project context becomes a real cost that does not shrink as the model price drops.
Who Should Use ZCode
ZCode makes sense for open-source projects, personal side projects, and developers evaluating open-weights models who do not handle sensitive code. The price advantage is real, the GLM-5.2 performance is competitive, and the Goal Mode workflow is a genuinely useful approach for long-horizon agentic tasks.
ZCode is the wrong tool for proprietary business logic, customer data, financial systems, healthcare applications, or any code subject to a data sovereignty review. The China National Intelligence Law exposure applies at the infrastructure level, not the model level — and it cannot be addressed by switching API keys.
The right question before installing is not “is GLM-5.2 good enough?” It is “is my codebase something I am comfortable routing through servers operating under a framework with no warrant requirement for government access?” For a significant portion of what developers build professionally, the honest answer is no. ZCode is a strong tool. Know what you are trading for the discount.
