Claude CLI

Claude Code — the claude CLI — can send its requests to any endpoint that speaks the Anthropic Messages API. Our /anthropic endpoint does, so pointing it at your workspace takes two environment variables and no code changes.

Once set, claude stops using your Anthropic subscription and bills against your Tensor Machine workspace instead, under your model allow-list, spend limits and audit trail.

1. Set the two variables

export ANTHROPIC_BASE_URL="https://edge.tensormachine.ai/<org>/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-tm-..."

Replace <org> with your workspace slug (the exact URL is on the API Keys page) and use a sk-tm-… key from that same page.

Two details that decide whether this works:

• No /v1 on the base URL. Claude Code appends /v1/messages itself. Ending the base URL at /anthropic is correct; adding the version yourself produces /anthropic/v1/v1/messages and a 404.
• ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. They send the credential in different headers — ANTHROPIC_AUTH_TOKEN as Authorization: Bearer, ANTHROPIC_API_KEY as x-api-key. Our endpoint accepts both, so either authenticates, but ANTHROPIC_AUTH_TOKEN takes effect immediately, while ANTHROPIC_API_KEY prompts you once in interactive mode to approve it over a saved login. Prefer the auth token.

2. Make it stick

Environment variables last only as long as the shell. To set them permanently, use Claude Code's user settings file:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://edge.tensormachine.ai/<org>/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "sk-tm-..."
  }
}

3. Choose a model

Behind a custom base URL, Claude Code passes the model name through without checking it — so pass a tm/… alias from My Team → Model access:

export ANTHROPIC_MODEL="tm/qwen3-5-9b"

Or set "model": "tm/..." at the top level of settings.json. Claude Code also runs a smaller model for background work like titling conversations; point that at an allowed model too, or those calls will request a model your workspace doesn't serve:

export ANTHROPIC_DEFAULT_HAIKU_MODEL="tm/qwen3-5-9b"

Model names you'd normally use with claude — sonnet, opus, haiku — are Anthropic's own and are not served here. Use your workspace's aliases.

4. Verify

Start claude from the same shell and run /status. Under Status you should see an Anthropic base URL line showing your Tensor Machine endpoint — that line only appears when a gateway address is set. If instead you see a Login method line naming a claude.ai account, the variables didn't take effect and requests are still going to Anthropic.

You can also check the endpoint directly, without involving the CLI:

curl "$ANTHROPIC_BASE_URL/v1/messages" \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"tm/qwen3-5-9b","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}'

A response starting with {"id":"msg_ means the endpoint, the key and the model alias are all good, and anything still broken is in the CLI's configuration.

Finally, confirm the request appears under Usage in the console — that's the proof it ran on your workspace rather than somewhere else.

Things worth knowing

• A saved Claude login stays saved. While the variables are set they take precedence; unset them and claude goes back to your normal account. Claude Code may warn at startup that both exist — that warning is expected here.
• Setting only the base URL is not enough. Without a credential variable, a saved claude.ai login remains the active credential and your requests won't reach us. Set both.
• Some features need Anthropic's own service and won't work against any third-party endpoint. If one is unavailable while you're pointed here, that's why.
• max_tokens is required by the Messages API. The CLI sets it for you; it only matters if you're testing with raw curl.

Troubleshooting

What we've verified

That our /anthropic endpoint accepts the Authorization: Bearer scheme Claude Code sends — not only the x-api-key header the Anthropic SDK uses — is covered by a test in our router, so this page's central claim is guarded against regression. The variable names and precedence rules above follow Anthropic's own documentation for connecting Claude Code to a gateway.

How well any given model performs as a coding agent is a separate question from whether the connection works; see Models.