⌘K
Guides

Claude Code and AI Hub

Point Claude Code at AI Hub through its Anthropic Messages API and use any catalogue model.

1. Prepare your workspace

Create a dedicated workspace inference key in this AI Hub environment and set a spend cap. Claude Code resends its system prompt and about 25 tool definitions on every turn, so a small task can use tens of thousands of input tokens. Choose exact author/slug IDs with tools support from Models or the catalogue:

bash
curl --fail "https://tokenhub-api.belive.sg/api/v1/models"

2. Configure Claude Code

Set these in the environment Claude Code runs in, from your secret manager. Leave ANTHROPIC_API_KEY empty so Claude Code uses the AI Hub key and not an Anthropic login:

bash
export ANTHROPIC_BASE_URL="https://tokenhub-api.belive.sg/api"
export ANTHROPIC_AUTH_TOKEN="$TOKENHUB_API_KEY"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen/qwen3.8-flash"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

Claude Code appends /v1/messages to ANTHROPIC_BASE_URL, so the base has no /v1. ANTHROPIC_DEFAULT_HAIKU_MODEL sets the model for Claude Code's background tasks; without it they use the main model. Use CLAUDE_CODE_SUBAGENT_MODEL to choose the subagent model.

3. Run

bash
claude --model qwen/qwen3.8-max

Claude Code may log that it does not recognize a model ID such as qwen/qwen3.8-max; it still sends it. Check Activity for the model, usage and cost of each turn.

What AI Hub supports

  • Messages on /v1/messages and /api/v1/messages, streamed or not, with tool use and tool results. Authorization: Bearer and x-api-key are both accepted.
  • cache_control on system and message blocks is passed to the model's provider; whether a provider caches is up to that provider.
  • Ignored rather than refused: extended thinking, context management, output effort, server tools such as web search, and thinking blocks in earlier turns. They are listed in the x-tokenhub-ignored-fields response header.
  • Refused with 400: mcp_servers, container and document (PDF) blocks. /v1/messages/count_tokens is not served; Claude Code then estimates token counts itself.

Every catalogue model with tools support works through this API, but how well each one follows Claude Code's tools varies by model.

Troubleshooting

  • authentication_error: the key is not in the environment Claude Code runs in, or a stored Anthropic login or ANTHROPIC_API_KEY is used instead. Run /status to see the active credential source.
  • not_found_error for the model: copy the exact catalogue ID, and check the key's model restrictions.
  • 404 for /v1/v1/messages: remove /v1 from the base.
  • For Claude Desktop, follow the Claude Desktop guide; shell variables are not its third-party inference settings.

Client support

For Claude Desktop, see the Claude Desktop guide.

← PreviousOpenCodeNext →Claude Desktop