Running GLM inside Claude Code

Point Claude Code at Z.ai’s GLM-5.2 to get a strong coding model at roughly a sixth of Opus token cost. Two shapes: delegate one subtask to GLM from inside an Opus session, or run a whole session on GLM. Your default claude stays on Opus either way.

GLM ships an Anthropic-compatible endpoint (https://api.z.ai/api/anthropic), so Claude Code talks to it with no adapter — just the standard ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_MODEL env vars.

Setup

Get a key from Z.ai, then add this block to ~/.zshrc:

# === GLM (z.AI) ===
export ZAI_API_KEY="your-zai-key-here"

# Run a WHOLE Claude Code session on GLM (default `claude` stays Opus):
claude-glm() {
  ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic" \
  ANTHROPIC_AUTH_TOKEN="$ZAI_API_KEY" \
  ANTHROPIC_MODEL="glm-5.2" \
  command claude "$@"
}

source ~/.zshrc, then claude-glm launches Claude Code on GLM-5.2. All normal flags pass through (claude-glm --resume, claude-glm -p "...").

Latest model on the key is glm-5.2 (glm-4.7 and glm-4.6 also work). Note: the “GLM-5.2 free on Cloudflare Workers AI” posts describe a different product — unrelated to a Z.ai key.

Delegating one task to GLM (the useful mode)

You usually don’t want to give up Opus. Instead, keep Opus driving and hand GLM individual self-contained subtasks — cheap grunt work — then use its answer.

Drop this helper at ~/.claude/scripts/glm.sh:

#!/bin/zsh
set -e
KEY="${ZAI_API_KEY:?ZAI_API_KEY not set}"
MODEL="glm-5.2"
if [[ "$1" == "-m" ]]; then MODEL="$2"; shift 2; fi
PROMPT="$*"
STDIN=""
if [[ -p /dev/stdin || -f /dev/stdin ]]; then STDIN="$(cat)"; fi
if [[ -n "$STDIN" ]]; then
  FULL=$'--- CONTEXT ---\n'"$STDIN"$'\n--- END CONTEXT ---\n\nTASK: '"$PROMPT"
else
  FULL="$PROMPT"
fi
printf '%s' "$FULL" | jq -Rs --arg m "$MODEL" \
  '{model:$m,max_tokens:4000,messages:[{role:"user",content:.}]}' \
  | curl -s -m 180 https://api.z.ai/api/anthropic/v1/messages \
      -H "x-api-key: $KEY" -H "anthropic-version: 2023-06-01" \
      -H "content-type: application/json" -d @- \
  | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["content"][0]["text"]) if "content" in d else sys.exit("GLM error: "+json.dumps(d))'

chmod +x ~/.claude/scripts/glm.sh, then use it:

# direct prompt (pass </dev/null in arg-only mode so it doesn't wait on stdin)
glm "write a Python function that merges overlapping intervals" < /dev/null

# pipe a file in as context
cat handler.py | glm "review for bugs, list max 5"

# model override
glm -m glm-4.7 "prompt" < /dev/null

Because it’s just a shell command, Claude Code (running Opus) can call it itself mid-session — say “delegate this to GLM” and it runs glm ..., then folds the result back into its work. GLM runs as a tool inside the session, not a replacement for it.

When to reach for GLM

Good fit

Isolated subtasks: boilerplate, one-off codegen, a quick review, a throwaway draft, format conversion.
Bulk or cost-sensitive work where Opus-grade judgement isn’t needed.
A cheap second opinion to compare against Opus.

Keep on Opus

Multi-step orchestration, or anything that leans on your hooks/agents/skills — GLM sees none of that, it’s raw prompt in, text out.
Correctness-critical single-shot work. In a small head-to-head, GLM-5.2 matched Sonnet 5 on straightforward coding (merge-intervals, a thread-safe LRU cache — both correct) but on a bug-diagnosis trap it reached the right insight and then hallucinated mid-answer. It also self-identifies as “Gemini.” Verify its output.

Limits worth knowing

You can’t route Claude Code’s native subagents (the Task tool / /agent) to GLM — those are Anthropic-model-only, no per-agent endpoint override. The shell helper is the only way to reach GLM from inside a session.
No automatic offloading. You (or Claude, with a stated reason) decide what to hand off.
Treat the key like any secret — env var, not committed.

Quick verification

glm "reply with exactly: GLM OK" < /dev/null     # -> GLM OK

If that returns GLM OK, the endpoint, key, and helper are all wired correctly.