2409324124 Claude Code Multi Provider Skill Claude Code Multi Provider

Claude Code Multi-Provider

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.

Known Entrypoints

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:

1. Opus / primary -> GPT-5.5 through raine-claude-code-proxy.
2. Sonnet -> DeepSeek deepseek-v4-pro.
3. Haiku -> MiMo mimo-v2.5-pro.
4. SubAgent and unmatched model names -> Gemini/Vertex proxy.

Important: a fallback launcher such as claude-auto may run cc-switch use mimo or cc-switch use deepseek, so it can change the default provider. For diagnosis, prefer individual entrypoints.

Router Architecture

The router (router.py) is a FastAPI proxy that sits between Claude Code and multiple LLM backends. It implements intelligent fallback inspired by LiteLLM.

Request Flow

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

Error Classification

The router classifies errors into two categories:

Per-Backend Configurable Retryable Statuses

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

Cooldown

Failed backends are placed on cooldown for COOLDOWN_SECONDS (default 300s). If the backend returns a Retry-After header, that value is used instead.

Fallback Chains

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

Health Endpoint

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]
}

Provider Shapes

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

Authentication Model

This setup uses mixed authentication. The router is not a vault for every provider credential:

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:

1. GPT: check raine-claude-code-proxy codex auth status, local port 18765, and the raine proxy log.
2. Gemini: check Google/Vertex credentials and local port 8083.
3. DeepSeek and MiMo: check router .env keys and endpoint URLs.
4. Router: check local port 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:

• works: 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-fast
• failed for the tested ChatGPT account: gpt-5.3-codex-spark-fast

DeepSeek 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.

Pitfalls

• Do not run 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.
• ChatGPT/Codex device auth and clawgate device auth can be separate flows. codex login --device-auth can succeed while clawgate login --default still waits for a different code.
• clawgate may not stay resident reliably with plain nohup. Use setsid ... >log 2>&1 < /dev/null &.
• clawgate ChatGPT mode warned that 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.
• GPT and Gemini usually do not need API keys in router .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.
• Do not use old DeepSeek deepseek-chat / deepseek-reasoner mappings for Claude Code if the official docs specify v4 Claude Code models.
• Do not copy terminal style artifacts like [1m] into model names. Treat them as ANSI formatting remnants unless the provider model list explicitly includes them.
• Gemini cannot be added to clawgate directly. Use a Gemini/Vertex Anthropic-compatible proxy or call Gemini CLI separately.
• A Gemini proxy must run from its repository directory. Starting uvicorn server:app elsewhere can fail with Could not import module "server".
• Some Gemini proxies only map Claude model names containing sonnet or haiku; set Claude defaults accordingly to trigger Gemini model mapping.
• Some Gemini proxy whitelists lag behind Google model releases. Add model IDs such as 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.
• Avoid fallback launchers during quiet audits if they mutate provider state.
• Claude Code settings can override shell-level ANTHROPIC_MODEL=... exports. For reliable route validation, update settings or use an entrypoint that fully controls the process environment.
• The router must strip incoming auth headers before forwarding. Do not pass authorization, x-api-key, anthropic-api-key, or anthropic-auth-token from Claude Code into arbitrary provider backends.
• Claude Code's native --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.

Safe Workflow

For status-only tasks:

1. Run the bundled diagnostic script.
2. Read masked provider config.
3. For GPT/Gemini, inspect local proxy auth and listener state before looking for router API keys.
4. For DeepSeek/MiMo, inspect router .env key presence and endpoint URLs.
5. Avoid running cc-switch use ... unless the user asked to switch providers.
6. Avoid 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'

Diagnostic Script

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.