How to Run GLM 5.2 in Claude Code, Pi, OpenCode & Every Harness (2026)

Matt Pocock (@mattpocockuk) asked: "Folks who are running GLM 5.2, how are you doing it? What harness/provider are you using?"

Update — July 19, 2026: Alibaba's Qwen Token Plan bundles GLM-5.2 + Qwen3.8-Max-Preview + DeepSeek V4 Pro on one subscription — same harness wiring as below, different base URL from qwencloud.com API Keys.

This guide answers in how-to form — one section per harness, step by step. Jump to the tool you use:

• How to run GLM 5.2 in Claude Code
• How to run GLM 5.2 in Pi
• How to run GLM 5.2 in ZCode
• How to run GLM 5.2 in OpenCode
• How to run GLM 5.2 in OpenClaw
• How to run GLM 5.2 in Codex
• How to run GLM 5.2 locally

Model context (benchmarks, vs Fable 5, export ban): GLM-5.2 explainer. Harness concepts: What Is an Agent Harness?.

Every section below follows the same pattern: what you need, which endpoint to use, how to select glm-5.2, and how to map reasoning effort. If you already have a Coding Plan key, jump straight to your harness. If not, start with How do I get a GLM Coding Plan API key?.

Quick pick: which harness?

How do I get a GLM Coding Plan API key?

Most how-tos below assume a GLM Coding Plan subscription (Lite ~$18/mo, Pro, Max, Team).

1. Sign up at z.ai
2. Open z.ai/manage-apikey/apikey-list
3. Create a Coding Plan API key (Team keys differ from personal keys — use the right one)

Endpoints you need:

Model IDs: glm-5.2 (standard) · glm-5.2[1m] (1M context in Claude Code)

Quota: GLM-5.2 costs 3× during peak (14:00–18:00 UTC+8), 2× off-peak — promo 1× off-peak through September 2026. Use GLM-4.7 for cheap routine tasks.

How to run GLM 5.2 in Claude Code

Z.ai lists Claude Code as a first-class Coding Plan client. You keep the same CLI; you swap the provider.

Step 1: Install Claude Code (if needed)

npm install -g @anthropic-ai/claude-code
cd your-project
claude

Step 2: Configure Z.ai credentials

Easiest — automated helper:

npx @z_ai/coding-helper

Manual — edit ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "your_zai_coding_plan_api_key",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "API_TIMEOUT_MS": "3000000"
  }
}

One-liner script (macOS/Linux):

curl -O "https://cdn.bigmodel.cn/install/claude_code_zai_env.sh" && bash ./claude_code_zai_env.sh

Restart the terminal or open a new Claude Code session after env changes.

Step 3: Select GLM-5.2

• Default model: GLM-5.2 or glm-5.2
• 1M context: glm-5.2[1m] and set CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000

Step 4: Set reasoning effort

Use /effort in Claude Code:

Use high for daily coding; max for architecture refactors and hard bugs.

Common Claude Code mistakes: using the general api/paas/v4 URL instead of the Anthropic-compatible path; putting the key in ANTHROPIC_API_KEY when Z.ai docs specify ANTHROPIC_AUTH_TOKEN; forgetting to restart the terminal after editing settings.json. If Claude Code still hits Anthropic's servers, run echo $ANTHROPIC_BASE_URL — it should print https://api.z.ai/api/anthropic.

Step 5: Run your first task

claude
# Example: fix a failing test with file context

More commands: Claude Code reference. Long sessions: loop engineering.

How to run GLM 5.2 in Pi

Tadej Stanic (@tadejstanic) on Matt's thread: "It's Opus 4.8 level for sure. When you use in pi.dev harness…"

Step 1: Install Pi

curl -fsSL https://pi.dev/install.sh | sh
# or: npm install -g --ignore-scripts @earendil-works/pi-coding-agent

Step 2: Start Pi in your project

cd your-project
pi

Step 3: Connect Z.AI Coding Plan

1. Run /login (OAuth) or export your Coding Plan API key per pi.dev/docs
2. Select provider Z.AI Coding Plan (not generic Z.AI if you are on the subscription)
3. /model or Ctrl+L → glm-5.2

Step 4: Configure reasoning effort

Pi maps effort for GLM-5.2 (earendil-works/pi#5770):

Step 5: Add verification (recommended)

Pi does not ship test loops by default. Paste a kickoff from explainx.ai loops — e.g. ci-until-green.

Deeper Pi context: Pi harness guide. Heavier fork: oh-my-pi with Z.AI/GLM in its provider matrix.

How to run GLM 5.2 in ZCode

Sheeki (@sheeki03): "zcode.z.ai — the app is a copy of Codex, and it's good as well."

Zero manual endpoint config — Z.ai's official desktop agent.

Step 1: Subscribe to GLM Coding Plan

z.ai/subscribe — GLM-5.2 included on all tiers.

Step 2: Download ZCode

Get the app from zcode.z.ai (macOS, Windows, Linux).

Step 3: Sign in and select GLM-5.2

1. Log in with your Z.ai account
2. Choose GLM-5.2 as the active model
3. Use /goal for long-horizon tasks (matches GLM-5.2's long-context positioning)

Check z.ai for in-app quota promos (community reported 50% extra usage and 1.5× effective quota offers — verify live before relying on them).

How to run GLM 5.2 in OpenCode

treysync (@0xtreysync): "OpenCode is my preference. Honestly been having comparable results to Opus."

Step 1: Install OpenCode

curl -fsSL https://opencode.ai/install | bash
# or: npm install -g opencode-ai

Full harness overview: OpenCode guide. Local open weights (no API): Run OSS models locally in OpenCode.

Step 2: Connect Z.AI Coding Plan

cd your-project
opencode

Inside the TUI:

1. /connect
2. Search Z.AI → select Z.AI Coding Plan
3. Paste your Coding Plan API key

Step 3: Select GLM-5.2

1. /models
2. Choose GLM-5.2 (use GLM-4.7 for routine work to save quota)

Step 4: Initialize project memory

/init

Creates AGENTS.md — same filename as Codex.

Slash commands: OpenCode commands reference.

Custom config (if needed in opencode.json):

{
  "provider": {
    "zai": {
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4"
      }
    }
  }
}

How to run GLM 5.2 in OpenClaw

OpenClaw is on Z.ai's supported tool list. It embeds Pi via SDK — configure GLM like any custom Anthropic provider.

Steps

1. Subscribe to GLM Coding Plan and get API key
2. In OpenClaw provider settings, set:
   • Base URL: https://api.z.ai/api/anthropic (Anthropic-compatible path)
   • API key: your Coding Plan key
   • Model: glm-5.2
3. Read Is OpenClaw Safe? before routing subscription OAuth through third-party harnesses

Underlying harness: Pi guide.

How to run GLM 5.2 in Codex (via Ollama)

For OpenAI Codex users who want GLM weights without Z.ai Coding Plan routing in Claude Code:

Step 1: Pull GLM-5.2 in Ollama

ollama pull glm-5.2

Step 2: Launch Codex in OSS mode

ollama launch codex
# or manually:
codex --oss -m glm-5.2

Full setup: Codex + Ollama OSS guide.

This uses Ollama as model host + Codex as harness — separate from Coding Plan quota rules.

How to run GLM 5.2 locally (LM Studio / vLLM)

Szymon (@Szymon_Lorenz) uses LM Studio. Henry Mascot noted ~256GB RAM for full local GLM-5.2 — plan for cluster/quantized serving, not a laptop default.

Step 1: Download weights

huggingface.co/zai-org/GLM-5.2 (MIT)

Step 2: Serve with your stack

• LM Studio — load model, enable OpenAI-compatible local server
• vLLM 0.23+ / SGLang 0.5.13+ — see GLM-5.1 local patterns (same workflow, model id glm-5.2)

Step 3: Point your harness at localhost

How to run GLM 5.2 in Cursor, Cline, and other tools

Z.ai documents these in quick start:

HCC (@HCColenbrander) warns Z.ai uptime can be spotty — keep GLM-4.7 or a fallback harness for deadline days.

How to run GLM 5.2 in Cursor specifically: open Settings → Models → Add custom provider. Set base URL to https://api.z.ai/api/coding/paas/v4, paste your Coding Plan key, and enter GLM-5.2 as the model name. Cursor treats this like any OpenAI-compatible endpoint — no Anthropic URL needed. Same pattern works in Windsurf, Continue, and other editors that accept a custom OpenAI base URL.

How to use Coding Plan MCP servers

Included on all tiers: Vision, Web Search, Web Reader, Zread (Z.ai docs). Enable in Claude Code or ZCode for mockup-to-code and live doc lookup.

How do I enable 1M context on GLM 5.2?

In Claude Code only (documented pattern):

1. Set model to glm-5.2[1m]
2. Export or set CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000
3. Use /effort max only when the task truly needs full-repo context

Other harnesses: check provider docs — OpenCode and ZCode may expose 1M natively depending on build.

How do I verify GLM 5.2 is actually running?

Before burning quota on a long agent session, confirm the model and endpoint are wired correctly:

1. Check the model name — the harness should show glm-5.2 or GLM-5.2, not a fallback like glm-4.7 or claude-sonnet.
2. Ask a identity probe — "What model are you?" GLM-5.2 should self-identify correctly; if it says Claude or GPT, your provider routing is wrong.
3. Watch the first API call — in Claude Code, enable verbose logging or check that requests go to api.z.ai, not api.anthropic.com.
4. Run a small coding task — fix one failing test or add a typed function. GLM-5.2 should produce coherent multi-file edits on the first turn.

If verification fails, re-check the base URL table in How do I get a GLM Coding Plan API key? — Anthropic-compatible harnesses and OpenAI-compatible harnesses use different paths.

How do I avoid burning through GLM 5.2 quota?

1. GLM-4.7 — lint fixes, small edits, doc Q&A
2. GLM-5.2 + high effort — features, multi-file refactors
3. GLM-5.2 + max effort — architecture, heisenbugs, /goal marathons
4. Schedule heavy runs outside 14:00–18:00 UTC+8 when off-peak promo applies
5. Add deterministic verification — agent loops so GLM is not the judge of "done"

Which harness should I pick?

None replace tests or CI as exit criteria — GLM-5.2 reasons well; your harness still needs a verification loop.

Summary

How to run GLM 5.2 is a harness question, not a model question. Pick your tool, follow the matching section above, and verify the endpoint before long sessions:

Model benchmarks and ban context: GLM-5.2 explainer.

Update — July 20, 2026: Hugging Face CEO said hosted US frontier models blocked exploit-payload analysis; HF switched to local GLM 5.2 for defensive security work. See cyber guardrails debate for IR routing patterns.

Related Reading

• AI cyber guardrails debate — self-hosted GLM 5.2 for IR (July 2026)
• Colibrì — GLM-5.2 on 25 GB RAM via disk streaming (July 2026)
• Perplexity Computer GLM 5.2 orchestrator — 0.344× Opus cost (July 9)
• Ollama $88M funding — what is Ollama? (July 2026)
• How to run open-source models locally in OpenCode — Ollama, llama.cpp, opencode.jsonc pillar guide
• Qwen 3.8-Max-Preview Token Plan — cloud frontier while waiting for 3.8 GGUF
• Qwen 3.6 27B local dev — llama.cpp + OpenCode on Apple Silicon
• GLM-5.2 Beats Fable 5 on Reasoning — China AI Response
• GPT-5.6 Sol in Claude Code — proxy setup like GLM harness swap
• What Is an Agent Harness?
• Top 10 AI Agent Loops for Coding Workflows
• Pi: Minimal Agent Harness by Mario Zechner
• GLM-5.1: Hugging Face and Ollama Run Guide
• Codex Open-Source Models via Ollama
• oh-my-pi: Pi Fork with Z.AI/GLM Support
• Z.ai GLM Coding Plan docs

Endpoints, quota multipliers, and supported tools are accurate as of June 19, 2026 per docs.z.ai. Verify current promos on z.ai before subscribing.