glm-launch

A Python CLI tool that wraps Claude Code with GLM settings. Instead of running a local proxy, it configures environment variables, then exec's the claude binary directly. (codex is not supported — Z.AI has no OpenAI Responses API endpoint.)

It works with Z.AI and their GLM series of models. You'll need a Z.AI API key — grab one with a Z.AI Coding Plan subscription. Using that referral link gives you 10% off and gets me 10% off too. Prefer not to? Here's a non-affiliate link.

Requires Python 3.13+.

Usage

# 1. Set your Z.AI auth token
export GLM_AUTH_TOKEN="your-zai-api-key"

# 2. Launch Claude Code routed through Z.AI (defaults to glm-5.3)
uv run glm-launch              # bare command defaults to `claude`
uv run glm-launch claude       # same thing, explicit

# Pick a different model
uv run glm-launch --model glm-5.3-flash            # bare options also go to `claude`
uv run glm-launch claude --model glm-5.3-flash     # multimodal, low cost, 3x quota
uv run glm-launch claude --model "glm-5.2[1m]"     # previous flagship, 1M tier
uv run glm-launch claude --model glm-4.5-air       # cheap

# Bootstrap your current shell so a plain `claude` uses Z.AI
eval "$(uv run glm-launch shell)"
claude

# See available models (built-in list, or --remote for the live API list)
uv run glm-launch models
uv run glm-launch models --remote

# Sanity-check connectivity / latency
uv run glm-launch bench

Examples use the installed glm-launch entrypoint. Before uv sync you can run the script directly with uv run src/main.py … — the two are interchangeable.

Installation

uv sync

This installs a glm-launch entrypoint. Run commands via uv run glm-launch <command>, or uv tool install . to get glm-launch on your PATH directly. You can also run the script without installing via uv run src/main.py <command>.

Run without cloning (uvx)

glm-launch is on PyPI, so you can run it directly with uvx (uv tool run) — no clone or manual install needed.

# From PyPI
uvx glm-launch launch claude

# Or straight from GitHub
uvx --from git+https://github.com/jefftriplett/glm-launch glm-launch launch claude

# Pin to a tag/branch/commit
uvx --from git+https://github.com/jefftriplett/glm-launch@main glm-launch models

Commands

launch claude

Launch Claude Code with GLM environment settings. Sets Anthropic env vars to route requests through Z.AI's Anthropic-compatible endpoint, then exec's the claude binary.

The launch prefix is optional: glm-launch claude is equivalent to glm-launch launch claude, and a bare glm-launch defaults to claude. Claude options can also be passed directly, so glm-launch --model glm-5.3-flash is equivalent to glm-launch claude --model glm-5.3-flash.

uv run glm-launch launch claude

Options:

The following env vars are set before exec'ing claude:

• ANTHROPIC_BASE_URL — from --base-url / GLM_BASE_URL
• ANTHROPIC_API_KEY — from --api-key / GLM_API_KEY
• ANTHROPIC_AUTH_TOKEN — from --auth-token / GLM_AUTH_TOKEN
• API_TIMEOUT_MS — from --api-timeout-ms / API_TIMEOUT_MS
• ANTHROPIC_DEFAULT_HAIKU_MODEL — from --default-haiku-model
• ANTHROPIC_DEFAULT_SONNET_MODEL — from --default-sonnet-model
• ANTHROPIC_DEFAULT_OPUS_MODEL — from --default-opus-model
• ANTHROPIC_DEFAULT_FABLE_MODEL — from --default-fable-model
• CLAUDE_CODE_SUBAGENT_MODEL — from --subagent-model
• CLAUDE_CODE_EFFORT_LEVEL — from --effort-level
• CLAUDE_CODE_ATTRIBUTION_HEADER — from --attribution-header
• CLAUDE_CODE_AUTO_COMPACT_WINDOW — from --auto-compact-window (only when non-empty)
• CLAUDE_CODE_MAX_CONTEXT_TOKENS — from --max-context-tokens (only when non-empty)

[!NOTE] With the default auto, the context settings are sized to the selected --model automatically: glm-5.3 and the [1m] IDs get 1M tokens, most other models 200K, and glm-4.5/glm-4.5-air 128K (unknown models fall back to 200K). glm-5.3 serves the 1M window natively; for other models the [1m] suffix is what enables Z.AI's 1M context tier — plain glm-5.2 and glm-5.3-flash serve the standard 200K window. Pass an explicit number to override, or an empty string to leave the env vars unset. Run glm-launch models to see each model's window.

Effort levels

Z.AI collapses Claude Code's effort ladder into the model's thinking tiers (source):

GLM-5.3 no longer supports disabling thinking entirely — low is the lightest setting. Z.AI recommends max for coding, which is the default here. You can also switch mid-session with the /effort command in Claude Code.

[!TIP] The GLM coding models are text-only — pasting images into Claude Code won't work through Z.AI. Coding Plan subscribers get image understanding via Z.AI's Vision MCP server (backed by glm-4.6v) instead; see #3. Also note that Team Plan API keys are separate from regular Z.AI keys — only a Team key draws Team quota, so a mismatched key can look like an auth failure.

Examples:

# Use defaults (glm-5.3 with 1M context, Z.AI endpoint)
uv run glm-launch launch claude

# The flagship (the default) — 1M context is standard on glm-5.3
uv run glm-launch launch claude --model glm-5.3

# Native multimodal (video/image/text/file) at a much lower cost,
# with 3x the coding-plan quota of glm-5.3
uv run glm-launch launch claude --model glm-5.3-flash

# Previous flagship with the 1M context tier (the coding plan
# auto-routes glm-5.2/glm-5.1 requests to glm-5.3)
uv run glm-launch launch claude --model "glm-5.2[1m]"

# Previous flagship on the standard 200K window (cheaper)
uv run glm-launch launch claude --model glm-5.2

# Balanced cost/performance coding model
uv run glm-launch launch claude --model glm-4.7

# Fast, speed-optimized GLM-5 variant
uv run glm-launch launch claude --model glm-5-turbo

# Lightweight, low-cost model for cheaper runs
uv run glm-launch launch claude --model glm-4.5-air

# Tune the model tiers independently (e.g. cheap subagents, flagship main)
uv run glm-launch launch claude \
  --model glm-5.3 \
  --subagent-model glm-4.5-air \
  --default-haiku-model glm-4.5-air

# Pass extra args through to claude
uv run glm-launch launch claude -- --verbose

# Inspect the command/env without launching claude
uv run glm-launch launch claude --dry-run

# Override via env vars
GLM_AUTH_TOKEN="my-token" uv run glm-launch launch claude

--dry-run does not require the claude binary to be installed.

Run uv run glm-launch models to see all valid model names (or --remote for the live list).

If claude is not on your PATH, the tool falls back to ~/.claude/local/claude.

launch codex (not supported)

Codex is not supported by glm-launch. Current codex only speaks the OpenAI Responses API (it removed wire_api = "chat"), but Z.AI's GLM endpoints are Anthropic Messages and OpenAI Chat Completions only — there is no /responses endpoint, so codex requests return 404. The codex command is intentionally disabled and exits with this explanation.

Use launch claude instead — it uses Z.AI's Anthropic-compatible endpoint. If Z.AI later ships a Responses-compatible endpoint, codex support can be revisited.

shell

Print export lines that bootstrap your current shell with the GLM env vars — without launching anything. Eval the output and a plain claude (or any Anthropic SDK tool) will talk to Z.AI.

eval "$(uv run glm-launch shell)"
claude

Accepts the same model/auth options as launch claude (--model, --auth-token, --default-*-model, etc.). Secrets are shell-quoted; empty values are skipped. Sets ANTHROPIC_MODEL plus all the ANTHROPIC_* / CLAUDE_CODE_* vars listed under launch claude.

# Inspect what would be exported
uv run glm-launch shell

# Bootstrap with a specific model
eval "$(uv run glm-launch shell --model glm-5.3-flash)"

models

List Z.AI GLM models. By default prints a built-in, annotated list; --remote fetches the live list from the Z.AI PaaS endpoint.

# Built-in list (no token needed)
uv run glm-launch models

# Live list from the API (needs GLM_AUTH_TOKEN)
uv run glm-launch models --remote

Options:

The live endpoint is the OpenAI-compatible coding PaaS base (/api/coding/paas/v4/models) and uses Authorization: Bearer <token> — distinct from the Anthropic-style chat base (/api/anthropic) used by launch claude and bench. Coding Plan keys only work through the coding endpoints; if you have a general Z.AI API key instead, point --models-url at https://api.z.ai/api/paas/v4/models.

bench

Time a single /v1/messages round-trip against the configured GLM endpoint. Useful as a sanity check that your auth token, base URL, and chosen model are reachable.

uv run glm-launch bench

Options:

Sends a minimal 32-token request and prints the round-trip time. Exits non-zero on HTTP error or timeout.

Example output:

  glm-5.3 via https://api.z.ai/api/anthropic
  OK (200) in 412ms

Verifying every model ID

glm-launch models prints the built-in registry, and models --remote lists what the API advertises — but neither proves a given ID is callable with your key. Z.ai rejects an unknown or unentitled model with HTTP 400 modelCode: does not exist, and the [1m] context IDs are a naming convention that never appears in the API's model list at all. bench --all is the check that actually calls each one:

uv run glm-launch bench --all

  probing 14 models via https://api.z.ai/api/anthropic

  ok   glm-5.3                200  1980ms
  FAIL glm-5.3-flash[1m]      400  427ms
  ok   glm-5.3-flash          200  1686ms
  SKIP glm-5v-turbo           429  752ms
  ...
  1 model(s) were rate limited and not verified: glm-5v-turbo

  2 of 14 model(s) failed to resolve.
  Rejected as `modelCode: does not exist`: glm-5.3-flash[1m], glm-5.2[1m]

A 429 is reported as SKIP, not a failure — it means the ID resolved but the key is out of quota, which says nothing about whether the model exists. Only IDs the API actually rejects count toward the non-zero exit.

The same probe is available as an opt-in test suite:

GLM_LIVE_TESTS=1 uv run pytest -m live -v

These are skipped by default (and in CI) since they need a real token and hit the network.

usage

Open the Z.AI usage/quota dashboard in your browser. Coding Plan quotas are tracked in 5-hour and weekly windows, and there is no API for quota data — the dashboard is the only place to see it.

uv run glm-launch usage

doctor

Check your environment for correct setup. Reports on environment variables and binary availability.

uv run glm-launch doctor

Checks performed:

• Authentication — Whether the required GLM_AUTH_TOKEN is set. The token is masked in output.
• Environment variables — Whether the optional GLM, Anthropic default-model, and Claude Code env vars used by the launch commands are set.
• Binaries — Whether claude is found on PATH (with fallback to ~/.claude/local/claude), including its version — [1m]-suffixed models need a recent Claude Code, so if claude reports the model doesn't exist, upgrade.

Exits with code 1 if GLM_AUTH_TOKEN is not set or the claude binary is missing, 0 otherwise.

Example output:

Environment variables:
  GLM_BASE_URL: (not set)
  GLM_API_KEY: (not set)
  GLM_AUTH_TOKEN: zai_***
  GLM_MODELS_URL: (not set)
  API_TIMEOUT_MS: (not set)
  ANTHROPIC_DEFAULT_HAIKU_MODEL: (not set)
  ANTHROPIC_DEFAULT_SONNET_MODEL: (not set)
  ANTHROPIC_DEFAULT_OPUS_MODEL: (not set)
  ANTHROPIC_DEFAULT_FABLE_MODEL: (not set)
  CLAUDE_CODE_SUBAGENT_MODEL: (not set)
  CLAUDE_CODE_EFFORT_LEVEL: (not set)
  CLAUDE_CODE_ATTRIBUTION_HEADER: (not set)
  CLAUDE_CODE_AUTO_COMPACT_WINDOW: (not set)
  CLAUDE_CODE_MAX_CONTEXT_TOKENS: (not set)

Binaries:
  claude: /usr/local/bin/claude

All checks passed.

Environment variables

All are optional except GLM_AUTH_TOKEN. glm-launch --help prints this same list, and glm-launch doctor shows which are currently set.

How it works

launch claude follows three steps:

1. Resolve the claude binary on PATH (falling back to ~/.claude/local/claude)
2. Set up the GLM environment variables
3. os.execvpe() the binary — fully replacing the glm process with claude for direct stdio passthrough

Z.AI exposes an Anthropic-compatible endpoint at https://api.z.ai/api/anthropic, so no local proxy is needed. The CLI sets the standard ANTHROPIC_* env vars and Claude Code talks directly to Z.AI.

Development

Common tasks are wrapped in a justfile. Run just with no arguments to list them.

Before committing, run the same core checks used by CI:

uv run pytest
uv tool run prek run --all-files
uv build

Versioning follows CalVer (YYYY.MM.INC1), and lint hooks (ruff, pyupgrade, validate-pyproject) are configured in .pre-commit-config.yaml and run with prek. CI runs the tests, lint checks, and package build; the release workflow reruns the tests before publishing.

Releases are automated. Run just release to bump the CalVer version, relock, and push the tag in one step. Pushing a YYYY.MM.INC1 tag triggers the GitHub Actions release workflow, which builds and publishes to PyPI via trusted publishing (OIDC, no API token). A plain git push never publishes — only the tag does.