Model access

n8v rolls new models out account by account. A model behind a rollout flag runs only for accounts that have the flag, and that holds however the work starts: the web and iOS apps, the CLI, the API, MCP, A2A, GitHub, Slack, Telegram and automations.

How it works, layer by layer

1. Rollout flags

Each gated model sits behind one or more flags, called its gates. Haiku 5.5 sits behind haiku55; Opus 5.5 High sits behind opus55 and opus47EffortLevels, and needs both. Models no flag controls, such as Sonnet 5, run for every account.

An account's value of a flag comes from the first of these that is set:

Early access (Settings, General, "Enable early access features") turns on some flags by default for the account, such as the Gemini agent and effort levels. An admin's value for everyone, or for the account, still comes first.

2. The gate

One rule decides, on the server: a model runs when every flag it needs is on for the account. The pickers in the apps list what the same rule allows, so a model a picker offers is never refused.

• A name n8v does not know is refused, including a provider's own id such as claude-opus-5-5. Use n8v's names, such as opus-5.5; the list has them all.
• Preferences only shape the pickers. Hiding a model in Settings, Agent never refuses it.
• The model a chat runs when none is named, your default, skips a model your account can no longer use, as the composer does. You can change it on the Subscription limits page. A default you chose in the Slack or Telegram settings is kept as you chose it.

3. Where it holds

A refused run never reaches the agent. A message sent through the API without a model runs the chat's last model; if that model's flag was turned off since, the API accepts the message and the chat ends in the refusal.

4. The record

Every change of who may use what is recorded, with who made it, from where, when, the value before and after, and the reason given:

A change an admin makes over the API or MCP is recorded the same way, as API or MCP, with the request id of the call that made it, so it can be found in the key's own activity too. Asking for the value a gate already has changes nothing and records nothing.

Changes are kept. Refusals are kept for 90 days, at most one a minute for the same account, model and surface.

See the models you can use

Every model n8v runs, in the order the apps list them, with whether your account can use it.

curl -sS https://n8v.online/api/v1/models \
  -H "Authorization: Bearer $N8V_API_KEY"

{
  "default": "sonnet-5",
  "models": [
    {
      "id": "sonnet-5",
      "name": "Sonnet 5",
      "agent": "claudeCode",
      "provider": "Claude Code",
      "enabled": true,
      "needs": null
    },
    {
      "id": "haiku-5.5",
      "name": "Haiku 5.5",
      "agent": "claudeCode",
      "provider": "Claude Code",
      "enabled": false,
      "needs": null
    }
  ]
}

Pass an id as `model` to start_chat or send_message. Without one, chats use sonnet-5.
Claude Code: sonnet-5 (Sonnet 5, default), opus-5.5 (Opus 5.5)
OpenAI Codex: gpt-5.5 (GPT-5.5, needs a ChatGPT sign-in)
1 more model is not enabled for this account; include_disabled lists it.

n8v models          # the models you can use, by provider
n8v models --all    # every model, with the ones not enabled for you

When a model is refused

{
  "error": {
    "code": "model_not_enabled",
    "message": "Model \"haiku-5.5\" is not enabled for this account",
    "param": "model",
    "hint": "GET /api/v1/models lists every model and whether this account can use it."
  }
}

{
  "error": {
    "code": "invalid_model",
    "message": "Unknown model \"claude-opus-5-5\". Did you mean \"opus-5.5\"?",
    "param": "model",
    "didYouMean": "opus-5.5",
    "hint": "GET /api/v1/models lists every model and whether this account can use it."
  }
}

{
  "lastError": {
    "code": "model_not_enabled",
    "message": "Haiku 5.5 Low is not enabled for this account.",
    "model": "haiku-5.5-low",
    "retry": "with_another_model"
  }
}

Error: Model "haiku-5.5" is not enabled for this account. `n8v models` lists the models you can use.
Error: Unknown model "claude-opus-5-5". Did you mean "opus-5.5"? `n8v models` lists the models you can use.

For admins: Model access

Admin, Model access shows every gate with what it opens, exactly who has it, and the week's refusals. A gate whose model sends prompts to an operator that keeps them is marked, with how many accounts could send it their code.

• Open a gate for its value for everyone (default, on or off) and how many accounts that reaches, the models it decides (Show details for each model's id, the other gates it needs and its refusals), the accounts with their own value, and its own history.
• Every change asks first, stating how many accounts have the gate now and after, and who will run each model: a model behind two gates opens only to accounts that have both. A change that takes access away is marked red. Give an optional reason; each change is recorded and can be undone from its notice, which puts back the value it replaced. Choosing the value a gate already has changes and records nothing, and says so.
• Turn on for an account, or, when a gate reaches everyone, Turn off for an account: an account's own value, which wins over the value for everyone.
• Check an account shows its value of every gate and where each comes from: its own value, the value for everyone, the default or early access.
• Activity lists changes to model gates and early access, and every refusal, as they happen; it reads Live while it refreshes itself, and Paused while older entries are open. From a refusal, check the account or turn on the gate it lacked.

The headline numbers filter the list: Restricted shows only gates some accounts lack, Refused, 7 days opens the refusals. Links can open the page ready: ?gate=haiku55 opens a gate, ?account=<id> checks an account, ?q=Haiku searches.

For admins: over the API and MCP

Everything on the page can be read and changed with an API key. Create it in Settings, API: admins see a group Model access, admins only with two scopes.

Only an admin can create such a key, and it expires within 90 days. Every call checks that its owner is still an admin and not banned: until then the key reads and changes nothing. No connected app can hold these scopes, so Claude or ChatGPT connected through Settings never can; over MCP they come only with such a key. Only flags that gate a model can be changed this way.

The key's activity in Settings, API lists each call without its answer, since answers name other accounts.

curl -sS -X PATCH https://n8v.online/api/v1/model-access/gates/haiku55 \
  -H "Authorization: Bearer $N8V_ADMIN_KEY" \
  -H "Content-Type: application/json" \
  -d '{"everyone": "on", "reason": "Haiku 5.5 for everyone at launch"}'

{ "flag": "haiku55", "everyone": "on", "was": "default", "changed": true }

Questions

Why does the model picker not show a model? It is not open to your account. Settings, Agent says so under the list, and what early access would add, if anything. Otherwise an admin decides.

Why did the API accept a message that the chat then refused? The message named no model, so it runs the chat's last model, and that model's flag was turned off after the chat began.

Does sending a message again change it? No. The message stays as you wrote it. The chat adds a marker that names the model it was sent with, and the agent receives the message unchanged. It carries on the chat as after a typed message: in the same session for a model of the same family, with the history replayed for another family or agent.

Can a provider's own model id get around a gate? No. n8v runs only the models it knows by its own names; any other name is refused.