Menu
Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.
Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
npx skills add https://github.com/tomevault-io/skills-registry --skill claude-code-multi-providerLa commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.
Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.
Basé sur la classification professionnelle SOC
Affichage de SKILL.md
| name | claude-code-multi-provider |
| description | >- Use when this capability is needed. |
Use this skill when the user asks about Claude Code provider setup, GPT/ChatGPT OAuth via raine/claude-code-proxy or clawgate, MiMo, DeepSeek, Gemini/Vertex, local tier routers, cc-switch, claude-router, claude-gpt, claude-gemini, or fallback launchers.
This skill is intentionally conservative: inspect first, avoid changing providers unless the user explicitly asks, and never print API keys or OAuth tokens.
Common entrypoints:
claude-router: Claude Code through a local tier router, default local proxy 127.0.0.1:8084.raine-claude-code-proxy: GPT/Codex OAuth backend, default local proxy 127.0.0.1:18765.claude-gpt: legacy Claude Code through clawgate, default local proxy 127.0.0.1:8082.claude-gemini: Claude Code through an Anthropic-compatible Gemini/Vertex proxy, default local proxy 127.0.0.1:8083.cc-switch use mimo && claude: MiMo Token Plan via Anthropic-compatible endpoint.cc-switch use deepseek && claude: DeepSeek official Claude Code integration.Final target routing:
raine-claude-code-proxy.deepseek-v4-pro.mimo-v2.5-pro.Important: a fallback launcher such as claude-auto may run or , so it can change the default provider. For diagnosis, prefer individual entrypoints.
cc-switch use mimocc-switch use deepseekThe router (router.py) is a FastAPI proxy that sits between Claude Code and multiple LLM backends. It implements intelligent fallback inspired by LiteLLM.
Claude Code → Router (:8084) → Backend
├─ Success → Return response
├─ Retryable error (429, 5xx, quota) → Cooldown + Try next fallback
├─ Fallback error (4xx) → Try next fallback (no cooldown)
└─ All backends exhausted → 502
The router classifies errors into two categories:
| Class | Behavior | Trigger |
|---|---|---|
retryable | Cooldown + fallback | 429, 502, 503, 504, per-backend extra codes, quota exhaustion in body |
fallback | Fallback only, no cooldown | 4xx (except 429) |
Some backends return non-standard status codes for quota exhaustion. For example, GPT (raine/claude-code-proxy) returns 400 for "no quota". Configure via .env:
# GPT returns 400 for "no quota" — treat as retryable
GPT_EXTRA_RETRYABLE=400
Failed backends are placed on cooldown for COOLDOWN_SECONDS (default 300s). If the backend returns a Retry-After header, that value is used instead.
Configure per-backend fallback order in .env:
GPT_FALLBACKS=mimo,deepseek,gemini
DEEPSEEK_FALLBACKS=mimo,gpt,gemini
MIMO_FALLBACKS=gemini,deepseek
GEMINI_FALLBACKS=deepseek,mimo
GET / returns backend status, stats, cooldown info, and configuration:
{
"status": "ok",
"backends": {
"gpt": {
"base_url": "http://127.0.0.1:18765",
"model": "gpt-5.5",
"on_cooldown": true,
"cooldown_remaining_s": 245,
"stats": {"success": 0, "fail": 1, "cooldown": 0},
"extra_retryable": [400]
}
},
"fallbacks": {"gpt": ["mimo", "deepseek", "gemini"]},
"global_retryable_statuses": [429, 502, 503, 504]
}
Router config:
ANTHROPIC_BASE_URL=http://127.0.0.1:8084
ANTHROPIC_MODEL=router/opus
ANTHROPIC_DEFAULT_OPUS_MODEL=router/opus
ANTHROPIC_DEFAULT_SONNET_MODEL=router/sonnet
ANTHROPIC_DEFAULT_HAIKU_MODEL=router/haiku
CLAUDE_CODE_SUBAGENT_MODEL=router/subagent
This setup uses mixed authentication. The router is not a vault for every provider credential:
| Hop | Auth mechanism | Stored where |
|---|---|---|
| Claude Code -> router | Local placeholder token such as ANTHROPIC_AUTH_TOKEN=local-router | ~/.claude/settings.json or wrapper env |
| router -> GPT | No API key in router; raine-claude-code-proxy performs Codex/ChatGPT device auth | ~/.config/claude-code-proxy/codex/auth.json |
| router -> Gemini | No API key in router; Gemini proxy performs Google/Vertex auth | service-account JSON, ADC, or Gemini OAuth files |
| router -> DeepSeek | Router injects DEEPSEEK_API_KEY | router .env |
| router -> MiMo | Router injects MIMO_API_KEY | router .env |
GPT_API_KEY and GEMINI_API_KEY are normally unset in this architecture. Do not copy OAuth refresh tokens, Google service-account JSON contents, or ChatGPT session data into the router .env.
Provider-specific diagnostic order:
raine-claude-code-proxy codex auth status, local port 18765, and the raine proxy log.8083..env keys and endpoint URLs.8084, health endpoint, fallback stats, and cooldowns.Route meanings:
router/opus -> gpt-5.5 through raine-claude-code-proxy.router/sonnet -> DeepSeek deepseek-v4-pro.router/haiku -> MiMo mimo-v2.5-pro.router/subagent and unmatched model names -> Gemini proxy.raine GPT/Codex proxy:
GPT_PROXY_PORT=18765
GPT_MODEL=gpt-5.5
Validated raine models:
gpt-5.5, gpt-5.5-fast, gpt-5.4, gpt-5.4-fast, gpt-5.3-codex, gpt-5.3-codex-fast, gpt-5.4-mini-fastgpt-5.3-codex-spark-fastDeepSeek official Claude Code config:
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
ANTHROPIC_MODEL=deepseek-v4-pro
ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
CLAUDE_CODE_EFFORT_LEVEL=max
MiMo Token Plan config:
ANTHROPIC_BASE_URL=https://token-plan-sgp.xiaomimimo.com/anthropic
ANTHROPIC_MODEL=mimo-v2.5-pro
ANTHROPIC_DEFAULT_OPUS_MODEL=mimo-v2.5-pro
ANTHROPIC_DEFAULT_SONNET_MODEL=mimo-v2.5-pro
ANTHROPIC_DEFAULT_HAIKU_MODEL=mimo-v2.5-pro
CLAUDE_CODE_SUBAGENT_MODEL=mimo-v2.5-pro
Gemini/Vertex proxy config:
GEMINI_PROXY_DIR=~/tools/claude-code-proxy-gemini
GEMINI_PROXY_PORT=8083
PREFERRED_PROVIDER=google
USE_VERTEX_AUTH=true
VERTEX_PROJECT=<your-gcp-project-id>
VERTEX_LOCATION=<your-vertex-location>
BIG_MODEL=gemini-3.1-pro-preview
SMALL_MODEL=gemini-3.1-flash-lite
Legacy clawgate GPT config:
CLAWGATE_PORT=8082
CLAWGATE_BIG_MODEL=gpt-5.4
CLAWGATE_MID_MODEL=gpt-5.3-codex
CLAWGATE_SMALL_MODEL=gpt-5.2-codex
Use clawgate as a fallback only. In the local integration, clawgate ChatGPT mode did not reliably satisfy the GPT-5.5 target.
curl | bash installers blindly. Inspect installer scripts, download binaries manually when possible, and verify checksums.clawgate --version may not exist. Use clawgate help, clawgate status, and clawgate account list.codex login --device-auth can succeed while clawgate login --default still waits for a different code.nohup. Use setsid ... >log 2>&1 < /dev/null &.gpt-5.5 was not in its known Codex model allowlist, and Opus requests through clawgate timed out under the 20 second acceptance window.raine/claude-code-proxy v0.0.13 successfully handled GPT-5.5 with Codex device auth in the tested setup..env; their upstream auth is handled by local proxy login/credential files.cc-switch status can show custom providers as Active: unknown. Check the URL and profile config.deepseek-chat / deepseek-reasoner mappings for Claude Code if the official docs specify v4 Claude Code models.[1m] into model names. Treat them as ANSI formatting remnants unless the provider model list explicitly includes them.uvicorn server:app elsewhere can fail with Could not import module "server".sonnet or haiku; set Claude defaults accordingly to trigger Gemini model mapping.gemini-3.1-pro-preview and gemini-3.1-flash-lite before setting them in .env.gcloud auth application-default login can fail in non-interactive shells at the verification code prompt. Existing GOOGLE_APPLICATION_CREDENTIALS service account JSON can be used where appropriate.ANTHROPIC_MODEL=... exports. For reliable route validation, update settings or use an entrypoint that fully controls the process environment.authorization, x-api-key, anthropic-api-key, or anthropic-auth-token from Claude Code into arbitrary provider backends.--fallback-model only works in --print mode and only handles overloaded (529) errors. It does NOT support cross-provider fallback (different API endpoints). The router is the correct solution for multi-provider fallback.For status-only tasks:
.env key presence and endpoint URLs.cc-switch use ... unless the user asked to switch providers.claude-auto unless the user accepts provider mutation.For non-mutating launcher health checks:
claude-router -p '只回复 OK'
claude-gpt -p '只回复 OK'
claude-gemini -p '只回复 OK'
For forced router route checks, prefer updating Claude Code settings or using a controlled wrapper. Shell-only overrides can be ignored when ~/.claude/settings.json has an env block.
For MiMo or DeepSeek health checks, tell the user that cc-switch use will change the current default provider before running:
cc-switch use mimo && claude -p '只回复 OK'
cc-switch use deepseek && claude -p '只回复 OK'
Run:
~/.codex/skills/claude-code-multi-provider/scripts/diagnose.sh
Optional non-mutating health checks for independent launchers:
~/.codex/skills/claude-code-multi-provider/scripts/diagnose.sh --health
--health checks claude-gpt and claude-gemini only. It intentionally does not check MiMo or DeepSeek because that would require cc-switch use, which changes the default provider.
Source: 2409324124/claude-code-multi-provider-skill — distributed by TomeVault.