Point NextChat at a custom OpenAI-compatible endpoint.

Quick answer: BASE_URL, one key, and a models list.

On a server deploy (Vercel, Docker, or bare Node), three environment variables do the whole job. BASE_URL overrides where API requests go; the README describes it as "override openai api request base url" with a default of https://api.openai.com, and the value is entered without /v1 because NextChat appends the request path itself. OPENAI_API_KEY carries the gateway key. CUSTOM_MODELS controls the model picker: plus adds a model, minus hides one, -all clears the default list, and name=displayName renames an entry. The detail that makes multi-vendor ids work is the provider suffix. NextChat ships separate client code paths for several vendors, so a bare claude id added to CUSTOM_MODELS can surface under the Anthropic path, which expects a different key and request format. Appending @OpenAI to the id, as in +claude-sonnet-4-6@OpenAI, pins the model to the OpenAI-compatible path so the request goes to your BASE_URL in standard chat-completions format regardless of which vendor trained the model.

BASE_URL=https://api.apisrouter.com        # no /v1
OPENAI_API_KEY=sk-APIsRouter-...
CUSTOM_MODELS=-all,+claude-sonnet-4-6@OpenAI,+gpt-5.5@OpenAI,+deepseek-v4-pro@OpenAI
DEFAULT_MODEL=claude-sonnet-4-6

How NextChat resolves its endpoint.

NextChat (ChatGPTNextWeb on GitHub, roughly 88K stars) is one of the most deployed chat frontends in existence: a light web app with one-click Vercel deployment plus packaged clients for iOS, macOS, Android, Windows, and Linux. Its popularity comes from exactly the mechanism this page uses, everything is a configuration surface, and the endpoint is just another setting. There are two of those surfaces. Server deploys read environment variables at build and boot: BASE_URL decides the host, OPENAI_API_KEY authenticates, and CUSTOM_MODELS shapes the picker for every user of that deployment. The clients and the web UI additionally expose an in-app path: Settings, Model Provider, pick OpenAI, then fill the endpoint and key fields and list extra ids in the custom model names field. The in-app path stores values locally per device, which makes it the right surface for a personal client, while env vars are the right surface for a deployment other people use. Either way, the request that leaves NextChat is a standard chat completion against your endpoint with the model id as a plain string. With a multi-vendor gateway behind BASE_URL, the same deployment serves Claude for long answers, a fast Gemini id for quick questions, and DeepSeek or GLM for bilingual traffic, all through the one key.

Full setup: server deploy or in-app settings.

For a Vercel deploy, set the variables in the project's environment settings and redeploy; Vercel bakes env values in at build, so editing a variable without redeploying changes nothing. For Docker, pass the same variables with -e flags or an env file. The CODE variable is worth setting on any public deployment, it gates access with a password so strangers cannot spend your key. The in-app path needs no deployment at all. In Settings, choose the OpenAI provider, replace the endpoint with https://api.apisrouter.com, paste the key, and add ids in the custom model names field using the same syntax as the env variable. This is how the desktop and mobile clients work with a gateway, and it is also the fastest way to test values before committing them to a server deploy. DEFAULT_MODEL picks what new chats start on, which matters more than it sounds on a shared deployment: most users never change the model, so the default id is where most of the spend lands. Set it deliberately to the model you want carrying casual traffic.

docker run -d -p 3000:3000 \
  -e BASE_URL=https://api.apisrouter.com \
  -e OPENAI_API_KEY=$APISROUTER_API_KEY \
  -e CUSTOM_MODELS="-all,+claude-sonnet-4-6@OpenAI,+claude-haiku-4-5-20251001@OpenAI,+gemini-3.5-flash@OpenAI" \
  -e DEFAULT_MODEL=claude-haiku-4-5-20251001 \
  -e CODE=your-access-password \
  yidadaa/chatgpt-next-web

Choosing models for the picker.

Because the whole picker bills through one key, the tuning loop is observational: run a week, read the per-model usage in the console, and reshape CUSTOM_MODELS around what people actually clicked rather than what you predicted.

• Start the list with -all. The stock picker is a long OpenAI-only menu; clearing it and adding four or five deliberate ids gives users a picker where every choice is one you priced.
• The default model carries the deployment. claude-haiku-4-5-20251001 or gemini-3.5-flash as DEFAULT_MODEL keeps the casual majority of turns fast and inexpensive, with stronger ids one click away.
• Long-form work earns a frontier slot. claude-sonnet-4-6 and gpt-5.5 are the picks users reach for when a draft or an analysis matters.
• Bilingual deployments should include deepseek-v4-pro or glm-5.2; NextChat has an enormous Chinese-speaking install base and those ids fit that traffic natively.
• Renames are free documentation: claude-sonnet-4-6=Sonnet (writing) style entries make the picker self-explanatory for non-technical users.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Failure modes specific to NextChat.

The /v1 mistake runs in the opposite direction from most tools. NextChat appends the request path to BASE_URL itself, so the value belongs without /v1; pasting https://api.apisrouter.com/v1 produces doubled paths that 404. Tools like LibreChat expect the /v1 included, which is exactly why people carry the wrong habit in both directions. A Claude id that errors about keys or headers is the missing @OpenAI suffix. Without it, NextChat can route the id through its native Anthropic path, which never consults BASE_URL for your gateway and expects vendor-format authentication. Pin every gateway id with @OpenAI and the requests all take the compatible path. Env edits that change nothing are a redeploy problem. On Vercel the variables are read at build; on Docker the container needs recreating. The in-app settings, by contrast, apply immediately but only on that device, which is the other half of the same confusion. CODE and OPENAI_API_KEY get swapped surprisingly often. CODE is the access password users type into the UI; the key is what the server spends. If users report an unauthorized page before any chat happens, that is CODE; if chats fail against the endpoint, that is the key.

Who routes NextChat through a gateway.

• People running a personal deployment on Vercel who want one metered key behind it instead of a subscription per vendor.
• Small teams sharing a NextChat instance, using CODE for access control and one gateway key so the usage log doubles as the cost report.
• Desktop and mobile client users pointing the in-app endpoint at a gateway to reach Claude, Gemini, and DeepSeek from one picker on every device.
• Bilingual users mixing GLM and DeepSeek ids with Claude and GPT in one deployment, which native provider silos make awkward.
• Developers without access to a given vendor's billing. Top-up based access with no card requirement removes the per-provider sign-up dependency.

Verify the endpoint and debug the first chat.

Prove the endpoint before deploying: list models with the key and confirm every id you plan to put in CUSTOM_MODELS appears, spelled exactly. The ids are forwarded as strings, so the /v1/models output is the authoritative spelling. Then deploy and send one message per model in the picker. A 404 on everything is the /v1-in-BASE_URL mistake. A 401 is the key, either wrong or not present in the environment the build actually used. An error on Claude ids only is the missing @OpenAI suffix. A picker showing models you never added means CUSTOM_MODELS lost its -all prefix or the variable did not reach the build. Once chats flow, the APIsRouter console shows per-request model, token counts, and spend. For a deployment with more than one user, that view answers the question every NextChat admin eventually asks, which model is quietly eating the balance, with data instead of guesses.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | head -50

FAQ