Connect a client to Model Gateway

Model Gateway speaks the OpenAI and Anthropic APIs. Any client built on either one moves to the gateway by changing two values: the base URL and the API key. Every call is then priced, counted against your budget, and visible in the Model Gateway studio.

Before you begin

An Atlan account with access to the workspace you will work in.
A client or SDK that speaks the OpenAI or Anthropic API.
The gateway base URLs:
Dialect Base URL
OpenAI https://api.atlan.com/openai/v1
Anthropic https://api.atlan.com/anthropic/v1

Dialect	Base URL
OpenAI	`https://api.atlan.com/openai/v1`
Anthropic	`https://api.atlan.com/anthropic/v1`

Get an API key

Create an agent

Open the Model Gateway studio, go to Agent access, and select Create agent. Enter a name, and for Workspace pick your personal workspace, the one with your name. Anyone can create an agent there. Creating one in a shared workspace takes admin on that workspace.

Copy the key

The API key is shown once, when the agent is created. Store it in your secret manager or shell environment as ATLAN_API_KEY. Never commit it to a repository. If you lose it, select Rotate key on the agent’s row to get a new one.

List the models you can call

curl https://api.atlan.com/openai/v1/models \
  -H "Authorization: Bearer $ATLAN_API_KEY"

The response lists every model you can route to, in OpenAI’s format. Each entry has an id and a modality, such as chat or embedding. Use the id of a chat model, for example global.anthropic.claude-haiku-5-5, as the model in the calls below.

Make your first call

curl
OpenAI SDK (Python)
Anthropic SDK (Python)
Claude Code
Codex

curl https://api.atlan.com/openai/v1/chat/completions \
  -H "Authorization: Bearer $ATLAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "<model-id>",
       "messages": [{"role": "user", "content": "Hello"}]}'

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://api.atlan.com/openai/v1",
    api_key=os.environ["ATLAN_API_KEY"],
)
reply = client.chat.completions.create(
    model="<model-id>",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)

The Anthropic SDK appends /v1/messages itself, so its base URL stops at /anthropic.

import os

from anthropic import Anthropic

client = Anthropic(
    base_url="https://api.atlan.com/anthropic",
    api_key=os.environ["ATLAN_API_KEY"],
)
reply = client.messages.create(
    model="<model-id>",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.content[0].text)

Point Claude Code at the Anthropic base URL, and name gateway model IDs for both its main model and its fast background model:

export ANTHROPIC_BASE_URL="https://api.atlan.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="$ATLAN_API_KEY"
export ANTHROPIC_MODEL="global.anthropic.claude-sonnet-5-5"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="global.anthropic.claude-haiku-5-5"
claude

Use any Claude id from GET /models for either variable.

Add the gateway as a model provider in ~/.codex/config.toml:

model = "global.anthropic.claude-sonnet-5-5"
model_provider = "atlan"
web_search = "disabled"

[model_providers.atlan]
name = "Atlan Model Gateway"
base_url = "https://api.atlan.com/openai/v1"
env_key = "ATLAN_API_KEY"
wire_api = "responses"

Then run codex with ATLAN_API_KEY set. Use any chat id from GET /models as model. Turn off Codex’s built-in web search, as above: the gateway does not run hosted tools.

The OpenAI dialect reaches every chat model, including Claude. Use the Anthropic dialect for Claude models.

Supported operations

Operation	Path	Notes
Chat completions	`POST /openai/v1/chat/completions`	Unary and streaming
Messages	`POST /anthropic/v1/messages`	Unary and streaming
Responses	`POST /openai/v1/responses`	Stateless only. Retrieve, delete, and cancel are not served.
Legacy completions	`POST /openai/v1/completions`	Answered by a chat model
Embeddings	`POST /openai/v1/embeddings`
List models	`GET /openai/v1/models`
Images, speech, transcription, moderation, rerank	`/openai/v1/...`	Same paths as the OpenAI API

Batches, assistants, vector stores, fine-tuning, and Anthropic’s token counting are not served and return 404. On the Anthropic dialect, metadata is accepted and ignored, and a context_management that edits the conversation is refused with 400.

Expected result

The call returns a model reply. Its tokens and cost are metered to the agent whose key made it.

Troubleshooting

Symptom	Cause and fix
`401`	The key is missing, mistyped, or revoked. Check that `ATLAN_API_KEY` is set in the shell that runs the client.
`404` with `not_found`	The `model` is not one you can route to, or the operation is not served. Copy the `id` exactly from `GET /models`, and check Supported operations.
`400` with `invalid_request`	The request does not fit the model, such as an embedding model on chat completions, or it uses a field the gateway refuses. Check the model’s `modality` in `GET /models`, and the fields under Supported operations.
`503` with `service_unavailable`	The model’s provider is unavailable. Retry later, or pick a model from another provider.
`429`	A budget or rate limit refused the call. The error names the limit that refused it. Ask your admin to raise it, or wait for the window to reset.
A model is missing from `GET /models`	Its provider or the model is disabled, or it is not in your workspace. Ask your admin to enable it.

Docs for AI agents

Connect a client to Model Gateway

Before you begin

Get an API key

List the models you can call

Make your first call

Supported operations

Expected result

Troubleshooting

Next steps

MCP Gateway

AI Gateway overview

​Before you begin

​Get an API key

​List the models you can call

​Make your first call

​Supported operations

​Expected result

​Troubleshooting

​Next steps

MCP Gateway

AI Gateway overview

Before you begin

Get an API key

List the models you can call

Make your first call

Supported operations

Expected result

Troubleshooting

Next steps