Run GLM-5.2 1M in Claude Code

This guide routes Claude Code's API calls through Z.ai GLM-5.2 1M instead of Anthropic -- same Claude Code UI, same skills, same workflow.

It uses Z.ai's Anthropic-compatible endpoint, so Claude Code can talk to GLM directly. No local proxy is required.

Most importantly: your normal Claude Code stays untouched. Plain claude keeps using your usual backend, such as Opus on Vertex AI. GLM only activates when you pass --bare --settings.

---

Setup

1. Get a Z.ai API key

1. Go to https://docs.z.ai
2. Sign in to Z.ai / Zhipu AI
3. Create an API key
4. Copy the key for the local setup below

2. Save the key locally

mkdir -p ~/.config/mg-glm && chmod 700 ~/.config/mg-glm
echo 'export ZAI_API_KEY="your-zai-key-here"' > ~/.config/mg-glm/key.env
chmod 600 ~/.config/mg-glm/key.env
source ~/.config/mg-glm/key.env

3. Create a Claude Code settings override

Most Claude Code installs already have a provider configured: Anthropic login, Vertex AI, AWS Bedrock, or an API key. Inline env-var prefixes can lose that precedence fight. The reliable approach is a dedicated --settings override file.

cat > ~/.config/mg-glm/claude-glm-settings.json <<EOF
{
  "env": {
    "CLAUDE_CODE_USE_VERTEX": "",
    "ANTHROPIC_VERTEX_PROJECT_ID": "",
    "CLOUD_ML_REGION": "",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "$ZAI_API_KEY",
    "ANTHROPIC_API_KEY": "$ZAI_API_KEY",
    "ANTHROPIC_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2[1m]",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "API_TIMEOUT_MS": "3000000"
  }
}
EOF
chmod 600 ~/.config/mg-glm/claude-glm-settings.json

The empty strings neutralize Vertex / Bedrock defaults for this one command. ANTHROPIC_BASE_URL points Claude Code at Z.ai, and the model keys map Claude Code's normal aliases to GLM.

4. Run Claude Code on GLM-5.2 1M

claude --bare \
  --settings ~/.config/mg-glm/claude-glm-settings.json \
  --model sonnet \
  "Build me a REST API with auth, JWT, and tests"

That command talks to GLM-5.2 1M. Plain claude without those flags keeps using your normal backend.

--bare is required. It tells Claude Code to read auth from the settings file instead of your OAuth keychain or auto-detected providers.

5. Use max effort for harder tasks

claude --bare \
  --settings ~/.config/mg-glm/claude-glm-settings.json \
  --model sonnet \
  --effort max \
  "Refactor this codebase and add tests"

Z.ai's Claude Code docs also mention /effort max and /effort ultracode for interactive sessions.

6. Verify the swap actually fired

claude --bare --settings ~/.config/mg-glm/claude-glm-settings.json \
  --debug-file /tmp/glm-check.log \
  --model sonnet \
  -p "Reply with: SWAP_TEST" < /dev/null

grep -E "api.z.ai|glm-5\\.2|/api/anthropic/v1/messages" /tmp/glm-check.log | head -5

You should see glm-5.2[1m], /api/anthropic/v1/messages, or api.z.ai in the debug output. If you see a Vertex URL like projects/.../publishers/anthropic, your normal backend is still winning and you should re-check --bare, --settings, and the empty Vertex keys in the JSON.

---

Run both backends side by side

Two terminals, two backends, no conflicts:

# Terminal 1 -- your normal Claude Code, unchanged
claude --model opus "Architect a multi-region failover system"

# Terminal 2 -- GLM-5.2 1M through Z.ai
claude --bare --settings ~/.config/mg-glm/claude-glm-settings.json \
  --model sonnet \
  "Refactor this codebase to async/await"

This is the whole trick: different settings sources, different backends, both alive.

Optional: shorten with an alias

# Add to your .zshrc or .bashrc
alias claude-glm='claude --bare --settings ~/.config/mg-glm/claude-glm-settings.json --model sonnet'

Then:

• claude-glm "Refactor this" -> GLM-5.2 1M
• claude "Hard problem" -> your default Claude Code backend

Do not alias plain claude unless you intentionally want GLM to become your default.

---

How it works

Claude Code respects ANTHROPIC_BASE_URL plus ANTHROPIC_AUTH_TOKEN / ANTHROPIC_API_KEY. Z.ai exposes an Anthropic-compatible endpoint at https://api.z.ai/api/anthropic, so this guide points Claude Code there and maps Claude's model aliases to GLM models.

The setup is isolated because the provider details live in ~/.config/mg-glm/claude-glm-settings.json, and you only use them when you pass --bare --settings. Your global ~/.claude/settings.json, shell profile, Vertex AI setup, and normal Opus workflow are not modified.

---

Troubleshooting

Auth conflict warning

Add --bare. It prevents Claude Code from mixing this key with managed login/keychain auth.

Still hitting Vertex or Bedrock

Run with --debug-file /tmp/glm-check.log, then inspect the URL:

grep -E "api.z.ai|publishers/anthropic|bedrock-runtime" /tmp/glm-check.log

If you see Vertex or Bedrock, re-check the settings JSON and make sure you launched with both --bare and --settings.

Do not put these exports in .zshrc

Keep the provider switch in the settings file. Global exports make it too easy to accidentally route normal Claude Code sessions away from your preferred backend.

---

Revert / uninstall

rm -rf ~/.config/mg-glm

Nothing else is required because this guide does not edit your global Claude Code settings.

---

Resources

• Z.ai docs: https://docs.z.ai
• Z.ai Claude Code guide: https://docs.z.ai/devpack/tool/claude
• Z.ai GLM-5.2 guide: https://docs.z.ai/guides/llm/glm-5.2
• Claude Code docs: https://docs.anthropic.com/en/docs/claude-code