From project key to a visible request.

1. Create an account and select a workspace/project.
2. Check published models and your available Coin. Top up using an active package when available.
3. Create a project API key. Copy it once and store it in your server's secret configuration.
4. Send a request to your endpoint with a published model.

# Replace YOUR_API_KEY with the key from the API keys page; never commit it or expose it in a frontend.
curl https://api.chanxai.my.id/v1/models \
  -H "Authorization: Bearer YOUR_API_KEY"

In PowerShell, use curl.exe and $env:CHANX_API_KEY.

Chanx AI speaks the OpenAI format and the Anthropic format, so most coding agents and apps work with one key. Pick your tool.

export ANTHROPIC_BASE_URL="https://api.chanxai.my.id"
export ANTHROPIC_AUTH_TOKEN="sk-chanx-YOUR_KEY"
export ANTHROPIC_MODEL="MODEL_SLUG"
export ANTHROPIC_DEFAULT_OPUS_MODEL="MODEL_SLUG"
export ANTHROPIC_DEFAULT_SONNET_MODEL="MODEL_SLUG"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="MODEL_SLUG"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="200000"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC="1"
claude

$env:ANTHROPIC_BASE_URL = "https://api.chanxai.my.id"
$env:ANTHROPIC_AUTH_TOKEN = "sk-chanx-YOUR_KEY"
$env:ANTHROPIC_MODEL = "MODEL_SLUG"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "MODEL_SLUG"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = "MODEL_SLUG"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "MODEL_SLUG"
$env:CLAUDE_CODE_MAX_CONTEXT_TOKENS = "200000"
$env:CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC = "1"
claude

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.chanxai.my.id",
    "ANTHROPIC_AUTH_TOKEN": "sk-chanx-YOUR_KEY",
    "ANTHROPIC_MODEL": "MODEL_SLUG",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MODEL_SLUG",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MODEL_SLUG",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MODEL_SLUG",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "200000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Public /v1 endpoints use Authorization: Bearer sk-chanx-…. The console uses a revocable cookie session; console mutations also require a CSRF token and an allowed Origin.

Restrict keys by project, model, IP/CIDR, expiry, rate and spending limits. Rename, disable, rotate or revoke them from API Keys. Rotation may allow an explicit overlap period. Revocation prevents subsequent authentication.

Provider and 9Router credentials stay on the server. Do not send a console cookie as an API key.

GET /v1/models returns models available to the authenticated key. The public catalog lists published configuration with prices; the console also shows upstream discovery. A discovered ID is not automatically available for inference.

Choose a model that supports the required context, output, tools and vision. The gateway rejects unsupported requests before reserving Coin whenever possible.

curl https://api.chanxai.my.id/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"YOUR_PUBLISHED_MODEL","messages":[{"role":"user","content":"Hello"}],"max_tokens":128}'

POST /v1/chat/completions accepts the documented Chat subset. Tool and vision support depend on the published model and validated request format. Consult the OpenAPI schema for accepted parameters; unsupported fields are not silently applied.

curl https://api.chanxai.my.id/v1/responses \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"YOUR_PUBLISHED_MODEL","instructions":"Be concise.","input":"Hello","max_output_tokens":128,"stream":false}'

The implemented subset supports text input, instructions, max_output_tokens and streaming. Tools, multimodal input, previous_response_id, stored conversation state and background mode are not supported. Requests containing unsupported options are rejected before reservation.

Set stream to true and use a client that consumes text/event-stream incrementally. Chat streams use data frames and a final [DONE]; Responses streams use typed events such as response.output_text.delta, response.completed and response.failed.

A successful terminal event is emitted only after verified usage and settlement are persisted. A failed or interrupted stream must not be interpreted as completed simply because some text arrived. Missing final usage enters reconciliation; emitted text is not authoritative token accounting.

Disconnects and ambiguous upstream failures may leave a reservation pending reconciliation. Inspect the request ID and billing state instead of repeatedly submitting the same expensive request.

Use an exact published slug to pin a model. SmartRoute aliases are smart-code, smart-fast, smart-value, smart-document and smart-premium. They require a published routing policy and eligible candidates.

Selections and fallback reasons are stored for inspection. Fallback respects capabilities, key restrictions, budget and price ceilings. The gateway does not silently downgrade a pinned model to bypass a budget.

Workspace, project and API-key policies support daily, monthly and per-request limits, protected balance, warning thresholds and hard limits. Daily/monthly windows use UTC. Exposure includes outstanding reservations so simultaneous requests cannot all spend the same remaining budget.

Redis controls short-lived rates and concurrency. PostgreSQL owns durable budgets and money. Admission fails closed when required abuse controls are unavailable.

1 Coin = 1,000,000 micro-Coin. Financial API values are integer strings. Do not convert them to JavaScript Number for arithmetic.

1. Estimate a range from request assumptions; the estimate is not guaranteed.
2. Reserve a conservative envelope from available balance.
3. Normalize verified input, output, cache read/write and reasoning usage into non-overlapping classes.
4. Apply the pinned pricing version, model/reasoning multipliers and margin. Round upward once in micro-Coin.
5. Settle verified actual cost and release unused reservation atomically.

Missing, malformed or ambiguous usage enters reconciliation. A partial stream is charged only for verified usage. Historical prices do not change when a new version is published.

QRIS top-ups credit only after an administrator verifies the payment. A redirect, screenshot or browser status cannot credit the wallet. Refunds and adjustments require compensating ledger entries; automatic top-up is not supported.

Keep the X-Chanx-Request-Id response header and the request ID in error metadata. Request Explorer shows status, route, verified usage and financial outcome without storing prompt contents by default.

Do not assume an AI request retry is free or deduplicated. Invoice creation uses an Idempotency-Key; retry the same intended invoice with the same key.