# AgentRouter — Overview

> Base Sepolia testnet. Wallet setup and spending require the owner's authorization. Follow one wallet path.

Discover the available models and use your purchased key with OpenAI- or Anthropic-compatible requests.

[Human guide](https://docs.agentrouter.cc/use/inference) · [Pricing](https://agentrouter.cc/pricing#agents) · [Documentation index](https://docs.agentrouter.cc/llms.txt)

## Using your key

The steps after purchasing a key are the same for both wallets. Keep the key private and follow one of the shared API guides.

- [Discover models](https://docs.agentrouter.cc/use/models) — read the public catalog.
- [OpenAI-compatible API](https://docs.agentrouter.cc/use/openai) — use the completion interface.
- [Anthropic-compatible API](https://docs.agentrouter.cc/use/anthropic) — use the messages interface.

Need a key first? Follow the [MetaMask](https://docs.agentrouter.cc/wallets/metamask) or [Coinbase](https://docs.agentrouter.cc/wallets/coinbase) wallet guide. For setup or request problems, see [Troubleshooting](https://docs.agentrouter.cc/troubleshooting).

## Discover models

# Read the public model catalog

No authentication or payment is required. HTTP 200 returns a JSON array,
not an OpenAI object/list envelope; an empty array is valid. Each model has
`id`, `protocols`, `context_length`, `max_output_tokens`, `input_per_1m`,
and `output_per_1m`. Prices are USD per **million** tokens as decimal
strings. Optional fields include aliases, cache prices, and `free`.

Use an offered model with the desired protocol; do not copy old fixture
prices. On HTTP 429, respect `Retry-After`. Browser cross-origin access
requires verified CORS.

```sh
curl -sS https://api.agentrouter.cc/v1/models
```

### Expected result

```json
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "HTTP 200 public model catalog body",
  "type": "array",
  "items": {
    "type": "object",
    "required": ["id", "input_per_1m", "output_per_1m", "context_length", "max_output_tokens", "protocols"],
    "properties": {
      "id": { "type": "string" },
      "aliases": { "type": "array", "items": { "type": "string" } },
      "input_per_1m": { "type": "string" },
      "output_per_1m": { "type": "string" },
      "cache_read_per_1m": { "type": "string" },
      "cache_write_per_1m": { "type": "string" },
      "free": { "type": "boolean" },
      "context_length": { "type": "integer" },
      "max_output_tokens": { "type": "integer" },
      "protocols": {
        "type": "array",
        "items": { "type": "string", "enum": ["openai", "anthropic"] }
      }
    }
  }
}
```

## OpenAI-compatible inference

# Run OpenAI-compatible inference with a purchased key

Verify `zai/glm-5.3` is offered with `openai` in the current catalog before
spending. Set `AGENTROUTER_KEY` privately. Create **one** identity per
intended request; retain it and the exact body for retries. Do not rerun
`uuidgen` to bypass an uncertain result.

Purchased keys do not support streaming. HTTP 200 returns an
OpenAI-compatible completion. Replays return request status, not the
original completion. HTTP 401 means an invalid key; 402 means allowance
exhausted. Do not create a fresh identity after an uncertain charge.

```sh
REQUEST_ID=$(uuidgen)

curl -sS https://api.agentrouter.cc/v1/chat/completions \
  -H "Authorization: Bearer ${AGENTROUTER_KEY:?set your key privately}" \
  -H "Idempotency-Key: $REQUEST_ID" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "zai/glm-5.3",
    "max_tokens": 512,
    "reasoning_effort": "low",
    "stream": false,
    "messages": [{ "role": "user", "content": "Reply with exactly OK." }]
  }'
```

### Expected result

# Expected inference result

HTTP 200 returns an OpenAI-compatible completion, with the answer in
`choices[0].message.content` and token counts in `usage`. The tested request returned
`OK`; model output can vary. A 200 status with no visible answer is not proof of a
useful reply. Keep the request identity if the result is uncertain; do not create a
new billable request automatically.

## Anthropic-compatible inference

# Run Anthropic-compatible inference with a purchased key

Verify `zai/glm-5.3` is offered with `anthropic` in the current catalog
before spending. Set `AGENTROUTER_KEY` privately. Retain this identity
and exact body for retries; generate another identity only for a genuinely
new intended request.

HTTP 200 returns an Anthropic-compatible messages object. Purchased keys
do not support SSE. Replays return request status, not the original
completion. Do not create a fresh identity after an uncertain charge.
The native `x-api-key` header is also supported.

```sh
REQUEST_ID=$(uuidgen)

curl -sS https://api.agentrouter.cc/v1/messages \
  -H "Authorization: Bearer ${AGENTROUTER_KEY:?set your key privately}" \
  -H "Idempotency-Key: $REQUEST_ID" \
  -H 'Content-Type: application/json' \
  -H 'anthropic-version: 2023-06-01' \
  -d '{
    "model": "zai/glm-5.3",
    "max_tokens": 512,
    "stream": false,
    "messages": [{ "role": "user", "content": "Reply with exactly OK." }]
  }'
```

### Expected result

# Expected inference result

HTTP 200 returns an Anthropic-compatible message with content and usage.
Output can stop at the requested token limit; a successful HTTP response
does not guarantee the requested phrase. Purchased keys use non-streaming responses.

Retain the request identity after a timeout or uncertain charge. A replay
returns request status, not the original completion; do not generate a new
identity merely to retry an uncertain result.