AnRouter Docs | OpenAI-Compatible API Guides

5-minute quickstart

Create an API key in the AnRouter dashboard, then copy a model ID from the models page. The examples below use the currently listed glm-5.3.

Python requires the official openai package. The Node.js example requires the openai package and must be saved as an ES module, for example example.mjs. Set ANROUTER_API_KEY in the process environment before running either example.

1 · Key

Create an API key

2 · URL

Use the correct Base URL

3 · Model

Copy the exact model ID

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ANROUTER_API_KEY"],
    base_url="https://api.anrouter.com/v1",
)

response = client.chat.completions.create(
    model="glm-5.3",
    messages=[{"role": "user", "content": "Introduce yourself in three sentences."}],
)
print(response.choices[0].message.content)

// Save as example.mjs, then run: node example.mjs
import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.ANROUTER_API_KEY,
  baseURL: "https://api.anrouter.com/v1",
});

const response = await client.chat.completions.create({
  model: "glm-5.3",
  messages: [{ role: "user", content: "Introduce yourself in three sentences." }],
});
console.log(response.choices[0].message.content);

curl https://api.anrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $ANROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3",
    "messages": [{"role": "user", "content": "Introduce yourself."}]
  }'

Success criteria: the response has HTTP status 200 and contains assistant output. Before production, separately test streaming, tool calling, and any required multimodal features. A successful text response does not prove every advanced feature is supported.

Codex CLI / desktop app

Codex CLI, IDE extensions, and Codex in the ChatGPT desktop app share the user-level ~/.codex/config.toml. Configure AnRouter as a separate provider. Do not edit auth.json to impersonate the built-in OpenAI login.

1. Install and verify

# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh

# Check
codex --version

Windows PowerShell: irm https://chatgpt.com/codex/install.ps1 | iex

2. Verify the Responses route first

curl https://api.anrouter.com/v1/responses \
  -H "Authorization: Bearer $ANROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3","input":"Reply with only: ok"}'

3. Add the user-level configuration

# ~/.codex/config.toml
model = "glm-5.3"
model_provider = "anrouter"

[model_providers.anrouter]
name = "AnRouter"
base_url = "https://api.anrouter.com/v1"
env_key = "ANROUTER_API_KEY"
wire_api = "responses"

The .codex folder and config.toml location in macOS Finder — macOS: open `~/.codex/` in Finder and edit `config.toml`. On Windows, use `%USERPROFILE%\.codex\config.toml`. Open the image at full size for detail.

Protocol requirement: responses is the only wire API supported for a Codex custom provider and is also the default when omitted. The selected AnRouter model and channel must work through POST /v1/responses.

4. Supply the key and launch

export ANROUTER_API_KEY="your_anrouter_api_key"
codex --model glm-5.3

# Open the desktop app
codex app

The desktop app or IDE extension may not inherit shell variables. Put ANROUTER_API_KEY=… in ~/.codex/.env, fully restart the app, and start a new session.

Do not hand-edit auth.json: auth.json and the built-in openai provider belong to the OpenAI sign-in flow. A separate AnRouter provider is clearer and does not overwrite an existing OpenAI login.

Reference: OpenAI Codex custom model providers

Claude Code / VS Code

Claude Code sends requests to an Anthropic Messages endpoint, so its Base URL must omit /v1. AnRouter accepts that protocol, but Anthropic does not promise full Claude Code compatibility for non-Claude models. Validate tool use and editing workflows before using one on a real project.

1. Install

# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash

claude --version
claude doctor

Windows PowerShell: irm https://claude.ai/install.ps1 | iex

2. Run a temporary test

export ANTHROPIC_BASE_URL="https://api.anrouter.com"
export ANTHROPIC_AUTH_TOKEN="your_anrouter_api_key"
export ANTHROPIC_MODEL="glm-5.3"

claude

3. Persist the Claude Code settings

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.anrouter.com",
    "ANTHROPIC_AUTH_TOKEN": "your_anrouter_api_key",
    "ANTHROPIC_MODEL": "glm-5.3"
  }
}

Model discovery limitation: Claude Code gateway discovery only adds IDs beginning with claude or anthropic. AnRouter IDs such as glm-5.3 do not appear automatically, so specify ANTHROPIC_MODEL explicitly.

Known endpoint limitation: this AnRouter deployment exposes POST /v1/messages but not Claude Code's optional POST /v1/messages/count_tokens route. Ordinary Messages requests can work, but a client feature that specifically calls server-side token counting may fail. Do not treat one successful chat as proof of full Claude Code compatibility.

4. VS Code extension

Install the official Claude Code extension and first confirm that claude can complete a request through AnRouter in a terminal.
The extension and CLI share ~/.claude/settings.json by default, so prefer the shared configuration above.
If the extension does not read the shared variables or still prompts for sign-in, press ⌘/Ctrl + Shift + P and run Preferences: Open User Settings (JSON).
After changing variables, run Developer: Reload Window and send a short test request.

Opening the VS Code user settings JSON from the Command Palette — Select **Preferences: Open User Settings (JSON)** in the Command Palette. The UI language does not change the command name. Open the image at full size for detail.

Only when extension-specific settings are needed, merge the following fields into the existing VS Code user settings without deleting other entries:

{
  "claudeCode.disableLoginPrompt": true,
  "claudeCode.environmentVariables": [
    { "name": "ANTHROPIC_BASE_URL", "value": "https://api.anrouter.com" },
    { "name": "ANTHROPIC_AUTH_TOKEN", "value": "your_anrouter_api_key" },
    { "name": "ANTHROPIC_MODEL", "value": "glm-5.3" }
  ]
}

Claude Code environment variable fields in VS Code settings.json — Use the screenshot to locate the fields; use the verified values immediately above rather than copying placeholders from the image.

5. Claude Desktop: the Code tab

Local Code sessions in Claude Desktop can reuse Claude Code settings. If the desktop app does not inherit shell variables, add them through the Local environment settings. This path still uses ANTHROPIC_BASE_URL=https://api.anrouter.com without /v1.

6. Claude Desktop: third-party inference

In Claude Desktop, open Help → Troubleshooting → Enable Developer Mode.
Open Developer → Configure Third-Party Inference… and choose Gateway.
Set Gateway base URL to https://api.anrouter.com, enter the AnRouter API key, choose bearer authentication, and select Static API key.
Test the connection, enter an exact AnRouter model ID in Models, then choose Apply locally.

Claude Desktop Configure third-party inference window — Set Gateway base URL to `https://api.anrouter.com` without `/v1`; Claude Desktop uses `/v1/messages` when it sends the request. Open the image at full size for detail.

Do not append /v1 in Claude clients: Claude Code, VS Code, and the Claude Desktop Gateway all use https://api.anrouter.com. Do not enter the complete /v1/messages request URL either. OpenAI-compatible clients such as Hermes and OpenCode use https://api.anrouter.com/v1. A read-only window indicates that the organization has locked these values through managed configuration; contact the administrator.

References: Claude Code LLM gateway · VS Code integration · Claude Desktop third-party inference

OpenCode

OpenCode can connect to Chat Completions through @ai-sdk/openai-compatible. Its TUI, desktop app, and IDE interfaces can reuse the same provider configuration.

1. Desktop app: add a custom provider

Open the provider settings in OpenCode Desktop and add a custom provider.
Set Provider ID to anrouter and the display name to AnRouter.
Set Base URL to https://api.anrouter.com/v1 and enter your AnRouter API key.

Add a model under Models. Its ID must exactly match the AnRouter models page or GET /v1/models, for example glm-5.3.
Leave request Headers empty unless a documented integration requires them, submit, then choose anrouter/glm-5.3 from the model picker.

OpenCode Desktop custom provider and model configuration in English

2. TUI / CLI: store credentials

Start OpenCode and enter /connect in the TUI.
Choose Other and enter anrouter as the provider ID.
Enter your AnRouter API key. This stores credentials only; the provider configuration below is still required.

3. Configure the provider manually

Image attachments in OpenCode: custom-provider models are treated as text-only unless image capability is declared. Keep both "attachment": true and "modalities": { "input": ["text", "image"], "output": ["text"] } on every AnRouter model you want to use with images. If these fields are omitted, OpenCode may display and resize an image but remove it before the API request, causing the model to say it cannot view images.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anrouter/glm-5.3",
  "provider": {
    "anrouter": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "AnRouter",
      "options": {
        "baseURL": "https://api.anrouter.com/v1"
      },
      "models": {
        "glm-5.3": {
          "name": "GLM 5.3",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "glm-5.3-flash": {
          "name": "GLM 5.3 Flash",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "deepseek-v4-flash-0731": {
          "name": "DeepSeek V4 Flash 0731",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "kimi-k2.6": {
          "name": "Kimi K2.6",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "minimax-m3": {
          "name": "MiniMax M3",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "qwen-3.8-27b": {
          "name": "Qwen3.8 27B",
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        }
      }
    }
  }
}

Put project configuration in opencode.json at the repository root, or personal global configuration in ~/.config/opencode/opencode.json. The example intentionally omits hard-coded context and output limits so client metadata cannot drift from the service's actual capabilities.

Restart required: after changing model capabilities, quit OpenCode completely and reopen it. Starting a new session or refreshing the page is not enough because OpenCode caches model capabilities at startup.

4. Verify

opencode auth list
opencode

# In the TUI
/models

Protocol choice: @ai-sdk/openai-compatible maps to /v1/chat/completions. Use @ai-sdk/openai only for a model/provider that is explicitly intended to use Responses.

Reference: OpenCode custom providers

Keys and authentication

OpenAI-compatible requests use a Bearer token. The AnRouter Anthropic Messages route accepts both the Bearer authentication used by Claude Code and the x-api-key header commonly used by Anthropic SDKs.

OpenAI-compatible / Responses

Authorization: Bearer YOUR_ANROUTER_API_KEY

Anthropic Messages (choose one)

Authorization: Bearer YOUR_ANROUTER_API_KEY
x-api-key: YOUR_ANROUTER_API_KEY

Keep keys in trusted environments only. Never put them in browser code, public repositories, screenshots, or committed project files. Revoke and replace a key immediately if exposure is suspected.

Endpoints and Base URLs

A Base URL is the prefix to which a client appends a route; it is not always the full request URL. Different clients append different paths, so one URL cannot be copied mechanically into every app.

Use case	Enter this	Final request
OpenAI SDK / OpenCode / OpenClaw / Hermes	https://api.anrouter.com/v1	/v1/chat/completions
Codex custom provider	https://api.anrouter.com/v1	/v1/responses
Claude Code / Claude Desktop Gateway / Anthropic SDK	https://api.anrouter.com	/v1/messages
Raw HTTP	Use the complete URL	https://api.anrouter.com/v1/…
Native Gemini REST	https://api.anrouter.com	/v1beta/models/{model}:generateContent

The most common mistake is a duplicate path. Putting the complete /v1/chat/completions endpoint into an SDK Base URL, or giving Claude Code https://api.anrouter.com/v1, can produce an invalid /v1/v1/… request.

Primary AnRouter API routes

Text protocols

POST /v1/chat/completions
POST /v1/responses
POST /v1/responses/compact
POST /v1/messages

Discovery and vectors

GET /v1/models
GET /v1/models/{model}
POST /v1/embeddings
POST /v1/rerank

Images and audio

POST /v1/images/generations
POST /v1/images/edits
POST /v1/audio/transcriptions
POST /v1/audio/speech

Native Gemini

POST /v1beta/models/{model}:generateContent
POST /v1beta/models/{model}:streamGenerateContent

A route existing does not mean every model supports that capability. The model, upstream channel, and request fields must all be compatible.

Protocols and request examples

OpenAI Chat Completions

curl https://api.anrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $ANROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3","messages":[{"role":"user","content":"Hello"}]}'

OpenAI Responses

curl https://api.anrouter.com/v1/responses \
  -H "Authorization: Bearer $ANROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3","input":"Review the edge cases in this function."}'

Anthropic Messages

curl https://api.anrouter.com/v1/messages \
  -H "x-api-key: $ANROUTER_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model":"glm-5.3",
    "max_tokens":1024,
    "messages":[{"role":"user","content":"Hello"}]
  }'

List current models

curl https://api.anrouter.com/v1/models \
  -H "Authorization: Bearer $ANROUTER_API_KEY"

Choosing a protocol: start ordinary third-party clients with Chat Completions; Codex requires Responses; Claude Code uses Anthropic Messages. Test tool calling, streaming, image input, and structured output against the actual model.

Troubleshooting

401 / Unauthorized

Check that the key is complete and has not been revoked. OpenAI-compatible requests require Authorization: Bearer; when Claude Code uses Bearer authentication, set ANTHROPIC_AUTH_TOKEN.

404 / Not Found

Inspect the final URL. Look for /v1/v1, a full endpoint entered as a Base URL, or a model ID that does not exactly match GET /v1/models.

429 / Too Many Requests

This can indicate token, concurrency, or upstream capacity limits. Use exponential backoff with jitter and a bounded retry count.

Chat works, but an agent fails

Agents usually also depend on tool calling, streaming, and a particular protocol. Test Codex on /v1/responses and Claude Code on /v1/messages.

When contacting support, include the time, request endpoint, model ID, HTTP status, and any request/error ID in the response. Never send the full API key; at most retain its first and last four characters for identification.

OpenClaw

An OpenClaw model provider key and a Gateway login token are different credentials. Use the AnRouter API key only for the model provider; do not put it in gateway.auth.token.

1. Install and check

curl -fsSL https://openclaw.ai/install.sh | bash
openclaw --version
openclaw doctor

2. Put the key in the global runtime environment

# ~/.openclaw/.env
ANROUTER_API_KEY=your_anrouter_api_key

3. Add the provider without replacing existing configuration

Run the additive command below on the machine that actually runs the Gateway. OpenClaw protects provider maps from destructive replacement; --merge preserves existing channels, tools, skills, providers, and Gateway settings.

Gateway location	Configuration to edit
Local macOS / Linux	~/.openclaw/openclaw.json
Native Windows	%USERPROFILE%\.openclaw\openclaw.json
Windows Hub / WSL	Edit `~/.openclaw/openclaw.json` inside the WSL environment that actually runs the Gateway
Remote Gateway	Sign in to the remote host and edit the remote user's configuration

Why there is no UI screenshot here: OpenClaw's entry point varies across local, WSL, Hub, and remote Gateways. The authoritative file is on the machine that actually runs the Gateway. Use the table above instead of editing the wrong device by following an unrelated screenshot.

openclaw config set models.providers.anrouter \
  '{"baseUrl":"https://api.anrouter.com/v1","apiKey":"${ANROUTER_API_KEY}","api":"openai-completions","models":[{"id":"glm-5.3","name":"GLM 5.3"},{"id":"glm-5.3-flash","name":"GLM 5.3 Flash"},{"id":"deepseek-v4-flash-0731","name":"DeepSeek V4 Flash 0731"},{"id":"kimi-k2.6","name":"Kimi K2.6"},{"id":"minimax-m3","name":"MiniMax M3"},{"id":"qwen-3.8-27b","name":"Qwen3.8 27B"}]}' \
  --strict-json --merge

4. Validate and restart the Gateway

openclaw config validate
openclaw models list
openclaw models set anrouter/glm-5.3
openclaw gateway restart
openclaw gateway status

Capability note: openai-completions is the Chat Completions adapter. OpenClaw is a tool-intensive agent, so validate tool calling, streaming, and long-context behavior—not only ordinary chat—before production.

References: OpenClaw custom providers · safe config set edits

Hermes Agent

The recommended setup uses hermes model to add a custom endpoint. Current Hermes documentation uses the top-level custom_providers list for named custom endpoints; the manual example below follows that documented schema.

1. Hermes Desktop visual setup

Open model/provider setup during first launch, or reopen it later from the model settings.
Choose Local / custom endpoint, set Endpoint to https://api.anrouter.com/v1, and enter the AnRouter API key.
Use an exact model ID such as glm-5.3. If API mode appears, choose chat_completions.
Save, start a new session, and send a short test message.

Hermes Desktop Local custom endpoint configured with the AnRouter endpoint — The screenshot now shows `https://api.anrouter.com/v1` in Endpoint. Enter your own AnRouter key on the next line and never publish a screenshot containing it. Open the image at full size for detail.

2. Install the CLI

# macOS / Linux / WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

hermes --version
hermes doctor

Windows PowerShell: iex (irm https://hermes-agent.nousresearch.com/install.ps1)

3. CLI recommended: interactive setup

hermes model

Provider	Custom endpoint
API Base URL	https://api.anrouter.com/v1
API Key	AnRouter API key
Model	glm-5.3
API Mode	chat_completions
Context length	Leave blank for endpoint discovery when possible; otherwise use the model catalog

4. Optional: named provider configuration

# ~/.hermes/config.yaml
custom_providers:
  - name: anrouter
    base_url: https://api.anrouter.com/v1
    key_env: ANROUTER_API_KEY
    api_mode: chat_completions

model:
  provider: custom:anrouter
  default: glm-5.3

# ~/.hermes/.env
ANROUTER_API_KEY=your_anrouter_api_key

5. Verify

Run hermes doctor to check the installation and configuration.
Start hermes and send a short prompt first.
Use /model custom:anrouter:glm-5.3 in a session. Exit and run hermes model to add or change a provider.

Context requirement: Hermes requires at least 64,000 tokens for agent tool workflows. Leave the wizard field blank when endpoint discovery is available; only add context_length manually when it matches the current AnRouter model catalog and actual endpoint behavior.

References: Hermes Desktop · Hermes custom endpoints

Other compatible clients

When a client explicitly supports an “OpenAI-compatible” or “Custom OpenAI endpoint,” the four values below will usually connect it. A similar product name does not prove that a custom Base URL is supported; do not force undocumented environment-variable overrides into software with no public configuration path.

Provider type	OpenAI compatible
Base URL	https://api.anrouter.com/v1
API key	AnRouter API key
Model	Copy from the models page or `GET /v1/models`

Confirm whether the final client request uses /v1/chat/completions or /v1/responses.
Start with a minimal text request, then test streaming, tools, vision, JSON, and other required capabilities.
If a client accepts only an official vendor key and exposes no Base URL field, do not claim that it natively supports AnRouter.

Security and production checklist

Key isolation

Create separate keys for people, CI, and production services. Do not share one long-lived key across a team. Use environment variables, a system secret manager, or the client's credential store.

Capability acceptance

Validate the exact model ID, protocol, streaming, tool calls, timeouts, and maximum inputs with realistic requests. Do not put an agent into production merely because ordinary chat succeeds.

Retry policy

Retry only recoverable failures, with exponential backoff, jitter, and a maximum attempt count. Log request IDs—not complete keys or sensitive prompts.

Change control

Clients and models change. After a client upgrade, protocol change, or model switch, rerun the minimal request and agent capability checks in a test environment.

AnRouter integration guide