DEV Community

Cover image for Running Claude Code on a model with fewer refusals (Anthropic-compatible API)
Majik
Majik

Posted on Originally published at docs.interodin.com

Running Claude Code on a model with fewer refusals (Anthropic-compatible API)

Claude Code reads its endpoint and credentials from environment variables, so you can point it at a different backend without patching anything. This is useful in security work, where a general model sometimes refuses routine tasks such as explaining a piece of malware from your own incident or writing a detection rule. This walkthrough runs Claude Code on Odin v1 and Odin Flash through InterOdin's API.

What you need

  • Claude Code installed
  • An InterOdin API key (starts with io-sk-), created in the console at https://interodin.com
  • A throwaway project, container or VM (see the warning below)

Option 1: a separate claudeodin command (recommended)

InterOdin publishes a setup script that adds a claudeodin command. It stores its settings in ~/.claude-odin, so your normal claude keeps working with its own account. The script and its exact usage are on the docs page: https://docs.interodin.com/en/integrations/claude-code

If you prefer to replace the normal command, the script has a --replace-claude flag. To undo everything, run it with delete. Each file is backed up first.

Option 2: manual setup

Add the variables to ~/.claude/settings.json, or export them in your shell:

  • ANTHROPIC_BASE_URL = https://api.interodin.com (no /v1 here)
  • ANTHROPIC_AUTH_TOKEN = your io-sk- key
  • ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL and CLAUDE_CODE_SUBAGENT_MODEL = odin-v1
  • ANTHROPIC_DEFAULT_HAIKU_MODEL = odin-flash
  • CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS = 1, which avoids 400 errors caused by request fields the API does not support
  • ENABLE_CLAUDEAI_MCP_SERVERS = false, so claude.ai connectors are not loaded

Map every model slot. If one is left on a default Claude name, the API returns 404 model_not_found.

Read this before running it: permissions

The setup script turns off permission prompts (bypassPermissions), so Claude Code can run commands and edit files without asking. Only use it in a throwaway project, container or VM. If you want the prompts back, run the script with --ask-permissions.

Limits to expect

  • odin-v1 is text-only. Images are replaced with a note.
  • PDFs and other documents fail with 400 unsupported_content.
  • Thinking is on or off only. There are no effort levels or budgets.
  • One visible turn can produce several billed API requests, so watch usage in the console.
  • Claude Code's WebSearch tool works; each search is billed on top of the tokens, at the per-search price on the pricing page.

Troubleshooting

  • 401: check the key.
  • 402: top up your balance. Billing is prepaid credits per token, with no subscription.
  • 404 model_not_found: map all model slots to odin-v1.
  • Login screen on first run: make sure the key is in a file Claude Code reads before first-run setup.
  • Conflicts with ANTHROPIC_API_KEY: run unset ANTHROPIC_API_KEY.

Notes on use

Fewer refusals does not mean no rules. InterOdin's Acceptable Use Policy forbids illegal and harmful uses, including creating or deploying malware and unauthorized access. Use it on systems and data you are authorised to work on.

The full, maintained version of this guide lives in the docs: https://docs.interodin.com/en/integrations/claude-code

Top comments (0)