Skip to main content

Claude Code

Claude Code reaches the Router over the Anthropic Messages API, so it needs configuration only — no plugin, no proxy.

Skip the manual setup

0gpc writes this configuration for you, previews every change, and can undo it.

Configure​

Create ~/.claude/settings.json (Windows: C:\Users\<username>\.claude\settings.json) and replace YOUR_API_KEY with a Router API key:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://router-api.0g.ai",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_API_KEY",
    "ANTHROPIC_API_KEY": "",

    "ANTHROPIC_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "0gm-1.0-35b-a3b",

    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "983616"
  },
  "modelOverrides": {
    "claude-sonnet-5": "0gm-1.0-35b-a3b"
  }
}

Then run:

claude --permission-mode auto
Private mode

Both models above are TEE-backed ("verifiability": "TeeML"). To guarantee every Claude Code request runs inside an enclave, create the key with trust mode Private in Dashboard → API Keys — that key routes only to TeeML providers regardless of what the client sends. See Privacy & ZDR.

Notes​

  • modelOverrides sets the model behind auto mode's permission gate, which runs before every non-read-only action. Use it rather than ANTHROPIC_DEFAULT_SONNET_MODEL: that one replaces the slot, leaving Claude Code unable to recognize the model, so it stops disabling reasoning and a yes/no check turns into a full reasoning pass.
  • ANTHROPIC_DEFAULT_HAIKU_MODEL covers background work — session names for claude --resume, status for commands like /usage.
  • CLAUDE_CODE_MAX_CONTEXT_TOKENS is the main model's real context window. Without it Claude Code assumes 200 K and compacts at a fifth of glm-5.3's 1 M.
  • --permission-mode auto is required per session, or set {"permissions": {"defaultMode": "auto"}} in ~/.claude/settings.json. With an API key, sessions otherwise start in Manual mode.
  • Everything under env also works as shell exports; modelOverrides does not.

Web search and web fetch​

  • WebFetch works with every model. Claude Code fetches the page on your machine and passes its text to the model.
  • WebSearch works only with the claude-* models. Claude Code sends Anthropic's web_search server tool, and the Router forwards it to the provider unchanged (see Web Search). The claude-* models run the search, billed per search. Other models, including glm-5.3 and 0gm-1.0-35b-a3b, ignore the tool without an error and answer without searching.

Claude Code can send the search in a separate request to the model in ANTHROPIC_DEFAULT_HAIKU_MODEL, not your main model. The config above sets that slot to 0gm-1.0-35b-a3b, so search can fail even with a claude-* main model. To use WebSearch, set that slot to a claude-* model as well. Background work then runs at that model's price.

For now, to search the web with other models, add a search MCP server and turn off the built-in tool so the model uses yours. Put this in .claude/settings.json (project) or ~/.claude/settings.json (user):

{
  "permissions": {
    "deny": ["WebSearch"]
  }
}

This turns off the built-in WebSearch for every model in that scope, including the claude-* models. Remove it to get the built-in search back.

Models​

Claude Code needs a model that serves the Anthropic Messages API:

curl -s https://router-api.0g.ai/v1/models \
  | jq -r '.data[] | select(.supported_formats | index("anthropic")) | .id'

Any of those work in any slot. See Models for pricing and context windows.

Last updated on