Routers - Pareto Inference

Choose a router below. Your app uses the router endpoint and key. The router uses your Pareto endpoint and key.

OmniRoute

Use the OmniRoute installation guide to start the router. Get a Pareto key and prepaid credits before you continue.

Add Pareto

Open Providers in the OmniRoute dashboard.
Select Add OpenAI Compatible.
Enter these values.

Field	Value
Name	`Pareto`
Prefix	`pareto`
API Type	`Chat Completions`
Base URL	`https://api.paretoinference.com/v1`

Keep the custom paths in Advanced Settings empty.
Select Add.
Open the new Pareto provider.
Select Add in Connections.
Enter a connection name.
Enter your Pareto key in API Key.
Enter z-ai/glm-5.3-flash in Default Model.
Select Save.
Add z-ai/glm-5.3-flash in the provider’s Models section if it is absent.

The API Key (for Check) field in the first form only checks the endpoint. It does not save a provider connection.

Connect your app

Create a router key in OmniRoute’s API Manager.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:20128/v1`
API key	Your OmniRoute key
Model	`pareto/z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. The pareto/ prefix selects the provider in OmniRoute. OmniRoute sends z-ai/glm-5.3-flash to Pareto.

Send a test request

Set OMNIROUTE_API_KEY to your router key in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:20128/v1/chat/completions \
  -H "Authorization: Bearer $OMNIROUTE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pareto/z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the provider key, base URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add fallback providers to an OmniRoute combo, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

9Router

Use the 9Router installation guide to start the router. Get a Pareto key and prepaid credits before you continue.

Add Pareto

Open Providers in the 9Router dashboard.
Select Add OpenAI Compatible under Custom Providers.
Enter these values.

Field	Value
Name	`Pareto`
Prefix	`pareto`
API Type	`Chat Completions`
Base URL	`https://api.paretoinference.com/v1`

Select Create.
Open the new Pareto provider.
Select Add API Key.
Enter a connection name.
Enter your Pareto key in API Key.
Enter z-ai/glm-5.3-flash in Default Model.
Select Save.
Add z-ai/glm-5.3-flash in the provider’s Models section if it is absent.

The Default Model field is required. The API Key (for Check) field in the first form does not save a provider connection.

Connect your app

Open Endpoint in the 9Router dashboard.
Select Create Key in API Keys.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:20128/v1`
API key	Your 9Router key
Model	`pareto/z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. The pareto/ prefix selects the provider in 9Router. 9Router sends z-ai/glm-5.3-flash to Pareto.

Send a test request

Set NINEROUTER_API_KEY to your router key in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:20128/v1/chat/completions \
  -H "Authorization: Bearer $NINEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pareto/z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the provider key, base URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add fallback providers to a 9Router combo, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

LiteLLM

Use the LiteLLM installation guide to install the proxy. Get a Pareto key and prepaid credits before you continue.

Add Pareto

Set PARETO_API_KEY to your Pareto key in the server environment.
Set LITELLM_MASTER_KEY to a separate secret that starts with sk-.
Save this configuration as config.yaml.

model_list:
  - model_name: pareto-flash
    litellm_params:
      model: openai/z-ai/glm-5.3-flash
      api_base: https://api.paretoinference.com/v1
      api_key: os.environ/PARETO_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

The openai/ prefix selects the OpenAI API format in LiteLLM. LiteLLM sends the model ID z-ai/glm-5.3-flash to Pareto. PARETO_API_KEY pays for Pareto usage. LITELLM_MASTER_KEY controls access to this local proxy. Keep both keys out of source control.

Start the proxy.

litellm --config config.yaml --host 127.0.0.1 --port 4000

If you already use LiteLLM, add the model entry to your existing configuration. Keep your existing proxy access settings.

Connect your app

Use these settings for the local proxy above.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://127.0.0.1:4000/v1`
API key	Your `LITELLM_MASTER_KEY` value
Model	`pareto-flash`

For a shared proxy, use a LiteLLM virtual key with access to pareto-flash. Keep the master key with the proxy administrator.

Send a test request

In a second shell, set LITELLM_API_KEY to the key that calls your proxy. Use the master key for the local setup above. Then run this request. It uses Pareto credits.

curl http://127.0.0.1:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pareto-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the provider key, base URL, and model mapping. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. You control any LiteLLM retry or fallback rules that you add. Other providers charge separately.

Bifrost

Use the Bifrost setup guide to install the gateway. Get a Pareto key and prepaid credits before you continue. Bifrost 2.2.2 passed chat, streaming, and tool-call tests on September 23, 2026.

Add Pareto

Set PARETO_API_KEY to your Pareto key in the server environment.
Save this configuration as config.json in the Bifrost app directory, such as ./data/config.json.

{
  "providers": {
    "pareto": {
      "keys": [
        {
          "name": "pareto",
          "value": "env.PARETO_API_KEY",
          "models": [
            "*"
          ],
          "weight": 1
        }
      ],
      "network_config": {
        "base_url": "https://api.paretoinference.com",
        "default_request_timeout_in_seconds": 300
      },
      "custom_provider_config": {
        "base_provider_type": "openai",
        "allowed_requests": {
          "chat_completion": true,
          "chat_completion_stream": true,
          "list_models": true
        }
      }
    }
  }
}

The base_url value has no /v1. Bifrost adds /v1/chat/completions. The env. prefix tells Bifrost to read the key from the environment.

Start Bifrost.

npx -y @maximhq/bifrost -app-dir ./data -port 8080

Connect your app

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:8080/v1`
Model	`pareto/z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. The pareto/ prefix selects the provider in Bifrost. Bifrost sends z-ai/glm-5.3-flash to Pareto. A local Bifrost without virtual keys does not require an API key from your app.

Send a test request

Run this request. It uses Pareto credits.

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pareto/z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the provider key, base URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add fallback providers in Bifrost, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

new-api

Use the new-api quick start to start new-api. Get a Pareto key and prepaid credits before you continue. new-api v1.0.0-rc.40 passed chat, streaming, and tool-call tests on September 23, 2026.

Add Pareto

Open Channels in the new-api console.
Select Add channel.
Enter these values.

Field	Value
Type	`OpenAI`
Name	`Pareto`
Base URL	`https://api.paretoinference.com`
Key	Your Pareto key
Models	`z-ai/glm-5.3-flash`
Group	`default`

Save the channel.
Select Test for the new channel.

The base URL has no /v1. new-api adds the API path. In self-use mode, new-api does not need model prices. In a multi-user deployment, set a model price or ratio for z-ai/glm-5.3-flash. Get the current prices from the model list.

Connect your app

Open Tokens in the new-api console.
Select Add token and copy the new key.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:3000/v1`
API key	Your new-api token
Model	`z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. Your app uses the new-api token, not your Pareto key.

Send a test request

Set NEWAPI_API_KEY to your new-api token in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:3000/v1/chat/completions \
  -H "Authorization: Bearer $NEWAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the channel key, base URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add other channels for this model, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

Portkey

Use the open-source Portkey AI Gateway. Get a Pareto key and prepaid credits before you continue. Gateway 1.15.2 passed chat, streaming, and tool-call tests from its Docker image on September 23, 2026.

Use the Docker image. In the same test, the npx @portkey-ai/gateway build returned HTTP 500 for every streaming request. This is a gateway issue, not a Pareto issue. See Portkey issue #1389.

Add Pareto

The gateway does not store provider settings. Your app sends your Pareto key and two headers with each request.

Start the gateway.

docker run -d --name portkey-gateway -p 127.0.0.1:8787:8787 portkeyai/gateway:latest

Connect your app

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:8787/v1`
API key	Your Pareto key
Header `x-portkey-provider`	`openai`
Header `x-portkey-custom-host`	`https://api.paretoinference.com/v1`
Model	`z-ai/glm-5.3-flash`

Your app must support custom headers. In the OpenAI SDKs, set them with default_headers (Python) or defaultHeaders (TypeScript).

Send a test request

Set PARETO_API_KEY to your Pareto key in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:8787/v1/chat/completions \
  -H "Authorization: Bearer $PARETO_API_KEY" \
  -H "Content-Type: application/json" \
  -H "x-portkey-provider: openai" \
  -H "x-portkey-custom-host: https://api.paretoinference.com/v1" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the key, both headers, and the model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. See Troubleshooting for Pareto errors.

CLIProxyAPI

Use the CLIProxyAPI guide to install the router. Get a Pareto key and prepaid credits before you continue. CLIProxyAPI v7.3.15 passed chat, streaming, and tool-call tests on September 23, 2026.

Add Pareto

Add this entry to the CLIProxyAPI config.yaml. Replace YOUR_PARETO_API_KEY with your Pareto key.

openai-compatibility:
  - name: "pareto"
    base-url: "https://api.paretoinference.com/v1"
    api-key-entries:
      - api-key: "YOUR_PARETO_API_KEY"
    models:
      - name: "z-ai/glm-5.3-flash"

Restart CLIProxyAPI.

The base URL includes /v1. CLIProxyAPI adds only /chat/completions. Without /v1, Pareto returns HTTP 404. If you set an alias for the model, your app must use the alias.

Connect your app

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:8317/v1`
API key	A key from `api-keys` in `config.yaml`
Model	`z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. Your app uses the CLIProxyAPI key, not your Pareto key.

Send a test request

Set CLIPROXY_API_KEY to a key from api-keys in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:8317/v1/chat/completions \
  -H "Authorization: Bearer $CLIPROXY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the Pareto key, the base URL, and the model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. See Troubleshooting for Pareto errors.

One API

Use the One API guide to start One API. Get a Pareto key and prepaid credits before you continue. One API v0.6.10 passed chat, streaming, and tool-call tests on September 23, 2026.

On Apple silicon and other ARM computers, use the justsong/one-api:v0.6.10 image. The latest image is for x86 only. The default One API console is in Chinese. The steps below give the English meaning of each label.

Add Pareto

Sign in to the One API console as root. Change the default password.
Open 渠道 (Channels).
Select 添加新的渠道 (Add new channel).
Enter these values.

Field	Value
类型 (Type)	`OpenAI`
名称 (Name)	`pareto`
分组 (Group)	`default`
模型 (Models)	`z-ai/glm-5.3-flash`
密钥 (Key)	Your Pareto key
代理 (Base URL)	`https://api.paretoinference.com`

Select 提交 (Submit).
Select 测试 (Test) for the new channel.

To add the model, type z-ai/glm-5.3-flash in the custom model field and select 填入 (Add). The base URL has no /v1. One API adds the API path.

Connect your app

Open 令牌 (Tokens).
Select 添加新的令牌 (Add new token) and copy the new key.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:3000/v1`
API key	Your One API token
Model	`z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. Your app uses the One API token, not your Pareto key.

Send a test request

Set ONE_API_KEY to your One API token in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:3000/v1/chat/completions \
  -H "Authorization: Bearer $ONE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

gpt-load

Use the gpt-load guide to start gpt-load 2. Get a Pareto key and prepaid credits before you continue. gpt-load v2.0.0-rc.31 passed chat, streaming, and tool-call tests on September 23, 2026.

Add Pareto

Sign in with the gpt-load management key.
Open Groups.
Select Add group.
Enter these values.

Field	Value
Channel	`OpenAI Compatible`
Upstream URL	`https://api.paretoinference.com/v1`
API keys	Your Pareto key
Models	`z-ai/glm-5.3-flash`
Name	`pareto`

Save the group.

To add the model, select Fetch models, or add z-ai/glm-5.3-flash manually. Keep the client alias empty. The upstream URL includes /v1. Without /v1, Pareto returns HTTP 404.

Connect your app

Open Access keys.
Select Create key. Give the key access to the pareto group.
Copy the new key.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:3001/v1`
API key	Your gpt-load access key
Model	`z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. Your app uses the gpt-load access key, not your Pareto key.

Send a test request

Set GPT_LOAD_API_KEY to your gpt-load access key in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer $GPT_LOAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the group key, upstream URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add other channels for this model, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

AxonHub

Use the AxonHub guide to start AxonHub. Get a Pareto key and prepaid credits before you continue. AxonHub v1.0.0-beta10 passed chat, streaming, and tool-call tests on September 23, 2026.

Add Pareto

Open Channels in the AxonHub console.
Select Add Channel.
Enter these values.

Field	Value
Provider	`OpenAI`
Channel Name	`pareto`
Base URL	`https://api.paretoinference.com/v1`
API Key	Your Pareto key
Models	`z-ai/glm-5.3-flash`
Test Model	`z-ai/glm-5.3-flash`

Save the channel.
Enable the channel. AxonHub creates new channels in the disabled state.
Select Test in the Enable Channel dialog.

Changes to a channel can take some seconds to apply.

If your app sends Anthropic Messages requests to AxonHub, AxonHub sends the Anthropic metadata field to Pareto. Pareto ignores this field and names it in the Pareto-Ignored-Parameters response header.

Connect your app

Open API Keys.
Select Create API Key and copy the new key.
Enter these settings in your app.

Setting	Value
API format	OpenAI Chat Completions
Base URL	`http://localhost:8090/v1`
API key	Your AxonHub API key
Model	`z-ai/glm-5.3-flash`

Use your router’s address if it runs elsewhere. Your app uses the AxonHub API key, not your Pareto key.

Send a test request

Set AXONHUB_API_KEY to your AxonHub API key in the shell environment. Then run this request. It uses Pareto credits.

curl http://localhost:8090/v1/chat/completions \
  -H "Authorization: Bearer $AXONHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3-flash",
    "messages": [{"role": "user", "content": "Say hello."}],
    "max_tokens": 128,
    "stream": true
  }'

The response streams data: events with a choices array. For an error, check the channel status, key, base URL, and model ID. Pareto returns HTTP 429 when the account has no credits for the request, when too many requests arrive at the same time, or when the model is at capacity (model_capacity, with a Retry-After header). Pareto returns HTTP 503 when it is temporarily unavailable, and HTTP 502 when a model request fails. If you add other channels for this model, you control those routes and pay those providers. See Troubleshooting for Pareto errors.

​OmniRoute

​Add Pareto

​Connect your app

​Send a test request

​9Router

​Add Pareto

​Connect your app

​Send a test request

​LiteLLM

​Add Pareto

​Connect your app

​Send a test request

​Bifrost

​Add Pareto

​Connect your app

​Send a test request

​new-api

​Add Pareto

​Connect your app

​Send a test request

​Portkey

​Add Pareto

​Connect your app

​Send a test request

​CLIProxyAPI

​Add Pareto

​Connect your app

​Send a test request

​One API

​Add Pareto

​Connect your app

​Send a test request

​gpt-load

​Add Pareto

​Connect your app

​Send a test request

​AxonHub

​Add Pareto

​Connect your app

​Send a test request

OmniRoute

Add Pareto

Connect your app

Send a test request

9Router

Add Pareto

Connect your app

Send a test request

LiteLLM

Add Pareto

Connect your app

Send a test request

Bifrost

Add Pareto

Connect your app

Send a test request

new-api

Add Pareto

Connect your app

Send a test request

Portkey

Add Pareto

Connect your app

Send a test request

CLIProxyAPI

Add Pareto

Connect your app

Send a test request

One API

Add Pareto

Connect your app

Send a test request

gpt-load

Add Pareto

Connect your app

Send a test request

AxonHub

Add Pareto

Connect your app

Send a test request