Quickstart

Tokand sits in front of your model API and shrinks what you send. Same model, same answer. You pay for less of the tool trace. If Tokand is down, the call goes through unchanged.

1. Point the client at Tokand

Keep your provider key on the server. The app sends a Tokand key and an OpenAI-style chat request.

curl http://localhost:8081/v1/chat/completions \
  -H "Authorization: Bearer $TOKAND_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "frontier",
    "tier": "measured",
    "messages": [
      {"role": "user", "content": "did the tests pass?"},
      {"role": "tool", "name": "pytest", "content": "Pytest: 12 passed"}
    ]
  }'

2. Read the savings on the response

Every response includes tokey.cost. X is the unfiltered prompt, Y is what the provider charged, Z is the Tokand fee. The dashboard sums those calls.

{
  "tokey": {
    "tier": "measured",
    "upstream_called": false,
    "cost": {
      "x_would_have_cost": 0.00126,
      "y_actual_provider_cost": 0,
      "z_tokey_fee": 0.000045,
      "saved": 0.001215
    }
  }
}

3. Open the dashboard

Local dev skips sign-in. Savings is at /dashboard. API keys, team, billing, and settings are in the sidebar.