provider-switch — switch LLM provider

How it works

Each alias sets 4-6 environment variables (ANTHROPIC_BASE_URL, auth key, model overrides) and runs claude. Claude Code reads these on startup and connects to the alternative provider instead of Anthropic.

Env vars only persist in the current shell session. Opening a new terminal resets everything — Claude Code uses your Anthropic subscription again.

Credential handling

Rule	Why
The API key is piped to the writer script on stdin, never passed as an argument	argv is visible to ps for the life of the process and is written verbatim into the session transcript. The script rejects a key given as an argument
The verify connectivity test sends the key through curl -s -K - (a config block on stdin), never an Authorization header on the command line — same rule for every provider, not only one	argv is world-readable via ps for the life of the process
The OpenRouter model-list lookup (custom model validation) sends no auth header at all	The catalogue endpoint is public and unauthenticated — there is nothing to send
There is no model-invisible channel for a runtime secret in Claude Code 2.1.233	AskUserQuestion returns plain text straight into the transcript, and plugin userConfig "sensitive": true is enable-time only — never substituted into skill or agent text. So the skill reads the key from an environment variable already set in your shell, or a file you create yourself; it never asks you to type one into the conversation
The ~/.zshrc.bak snapshot is created under umask 077, chmod 600, and deleted once the write succeeds	The backup holds whatever keys were already in ~/.zshrc; a world-readable copy nothing cleans up is a second leak
Line edits go through awk + temp file + mv	sed -i '' is BSD-only, so set-key / set-alias / remove-key / remove-alias were broken on every Linux machine
DEEPSEEK_API_KEY is part of the status key list	It was missing, so DeepSeek — the priority default — re-prompted for a key it already had on every run

Auth by provider

Provider	Auth pattern	Notes
All providers	ANTHROPIC_AUTH_TOKEN + ANTHROPIC_API_KEY=""	Unified: Bearer token + empty API key blocks OAuth fallback

Context window

Claude Code defaults to 200K for non-Anthropic providers. The [1m] suffix on model names (e.g., qwen3.7-plus[1m]) forces 1M context. Applied automatically by the skill.

Skill arguments

Argument	What it does
(none)	Show status table
install	Interactive setup (was setup before v5.0.0)
deepseek / ds / glm / qwen / minimax / openrouter	Single-provider setup
verify	Test all configured tokens against endpoints
model-check	Identify which model is actually responding (run inside provider session). Matched before status — otherwise the status|check pattern swallows it as a substring
help	Switching cheat sheet

Research delegation

When the skill refreshes its provider reference files, it spawns one agent per provider — five Task calls in a single message (DeepSeek, Z.ai/GLM, Qwen/DashScope, MiniMax, OpenRouter). Task is in the skill’s allowed-tools, so that fan-out actually happens instead of collapsing into a serial single-agent crawl.

A big task handed to one agent is an agent gone for an hour: you cannot observe it, cannot correct it, and it usually drifts off-target. One subagent = one provider, ~5 files, ~10 steps — never all five providers in one prompt.

Each research prompt carries six fields; a bare one-line task is never enough:

Field	Content
GOAL	refresh the reference files so users get current model ids, pricing and endpoints
ROLE	research and report only — the skill applies changes later, the agent edits nothing
SCOPE	in — that provider’s doc sources; out — other providers, ~/.zshrc, any reference file on disk
CONTEXT	the current on-disk baseline (model, alias, endpoint) to diff against; four siblings fetch the rest
CONSUMER	the skill diffs the table, shows it to the user, who approves each update before anything is written
DONE	a Field / Current / Fetched / Source URL table — say “no change” explicitly, never omit a field

An omitted field reads as “unknown” downstream and blocks that provider’s whole update, which is why DONE spells out the shape.

Claude Code automatically sends cache_control markers to reduce input token costs. Only input tokens can be cached — output tokens are always billed at full rate, regardless of provider.

Provider support

Provider	Prompt Cache	Status	Details
DeepSeek V4	Yes	Verified	Automatic on 1024+ token repeated prefixes. 90% discount ($0.03/M on Pro). No opt-in required
MiniMax	Yes	Verified	Fully documented — pricing, response fields, TTL
Qwen	Yes	Verified	Implicit cache 20%, explicit 10% of standard rate. Min 1024 tokens. Confirmed in Alibaba billing
Z.ai / GLM	Yes	Verified	Backend auto-cache (native). Applied transparently at the provider side
OpenRouter	Model-dependent	—	Anthropic models: yes. Non-Anthropic (Qwen, GLM): routed to provider cache

MiniMax cache pricing (MiniMax-M3)

Token type	Price / 1M	vs Standard input
Standard input	$0.30	—
Cache write (first request)	$0.375	+25%
Cache read (subsequent)	$0.06	-80%

Cache TTL: 5 minutes, auto-refreshed on each hit. Minimum: 512 input tokens.

What this means in practice

A typical Claude Code request (~250K input tokens) with 90% cache hit rate:

Provider	Cost per request	Without cache
DeepSeek V4 Pro (cached)	~$0.01	$0.44
MiniMax (cached)	~$0.02	$0.08
Qwen (cached, explicit 10%)	~$0.02	$0.13
Anthropic Opus 4.8 (cached)	~$0.24	$1.25
Z.ai / GLM (backend cache)	~$0.35	$0.35