Models and providers

Choose a modelCopy Anchor Link

Pick a model in the composer’s model menu. Models with reasoning levels have a reasoning submenu:

• None
• Low
• Medium
• High
• Extra High
• Max

Each model marks one level as its default. Service-tier switches appear below the reasoning choices. Turn the active switch off to use Standard speed. Sprocket saves the model, reasoning level, and service tier for each thread.

The model menu shows the gateway catalog’s models and options, whether or not you have connected a provider for them. Requests follow your provider order. If an option requires your own provider, the failed request explains what to connect in Settings.

UltrafastCopy Anchor Link

GPT-6.1 Sol offers an Ultrafast switch below its reasoning choices. Connect your own OpenAI API key and place OpenAI API before Spikonado in Settings → Providers to use it. Turning on Fast or Ultrafast turns the other off.

Ultrafast uses your provider account, not your Sprocket plan. OpenAI charges six times Sol’s Standard API prices. Spikonado rejects Ultrafast requests with a message asking you to connect your own provider. Changing provider settings does not hide the switch or reset your selection.

From the CLI:

sprocket run --model gpt-6.1-sol --service-tier ultrafast "Investigate the failing tests"

Sprocket loads the model list from its servers, so new models show up without an update. To see the list from your terminal, run:

sprocket run --list-models

ProvidersCopy Anchor Link

A provider is where Sprocket sends model requests. Spikonado, Sprocket’s own provider, is on by default and uses your plan’s AI usage. The other providers use your own keys or subscriptions and don’t count against your Sprocket AI usage.

Add keys and subscriptions in Settings → BYOK/BYOS. Sprocket checks each API key with its provider before saving it.

Claude and Muse sign-in run on the computer that runs Sprocket. Open Sprocket on that computer to install them or sign in.

MorphCopy Anchor Link

Morph serves Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017. Pay-as-you-go keys and Ultra plan keys connect the same way. On Ultra, neither model has a token meter, but each has a per-minute limit on uncached input and output tokens. Kimi K3.1 Pro’s output limit is lower than one nonstop stream, so threads with several agents hit it quickly. Keep another provider enabled after Morph so those requests fall back instead of failing.

Provider order and fallbackCopy Anchor Link

In Settings → Providers, drag providers into the order you want and turn each on or off. At least one provider must stay on.

Connecting a provider in Settings → BYOK/BYOS adds it to the end of the list. Removing a provider from either page takes it off the list and deletes what Sprocket saved for it: an API key, the Claude token, or every ChatGPT sign-in. Spikonado and Muse can’t be removed, since Sprocket holds no credential for them. Turn them off instead. If you remove the last provider that’s on, Sprocket turns Spikonado back on.

For each request, Sprocket uses the first enabled provider that offers the model. Sprocket retries server errors. If the provider still fails or rate-limits the request, Sprocket moves the run to the next enabled provider. If no enabled provider offers the model, the run fails with an error that points to Settings → Providers.

Long threadsCopy Anchor Link

When a thread’s context gets close to the model’s limit, the model writes a handoff note covering the task, its progress, and what’s left. Sprocket continues the thread in a fresh context that starts from that note. It does the same when a thread switches to another provider or to a model from a different vendor.