Providers and models
Studio ships no model of its own. You connect a provider, then pick which model this session calls from the composer. The model a new session starts on is the default under Sessions.
Settings → Providers (keys and endpoints) and the model list on each provider tile. Reasoning effort and Fast mode live on Settings → Model.
API key
Add a key from the Providers page. Pick the provider, paste the key.
Subscription login
Where supported, sign in with a consumer plan instead of pasting a key (for example ChatGPT, GitHub Copilot, SuperGrok, OpenRouter, Chutes). Browser or device-code flows store a token beside your keys.
Self-hosted providers ProTeamEnterprise
Attaching a local runtime (Ollama, LM Studio, llama.cpp, LocalAI, vLLM, ...) or defining a custom OpenAI-compatible base URL. A local connection is edited in one place, and the card shows whether the last probe reached it.
Vertex AI, Amazon Bedrock, and Azure OpenAI use their own auth (ADC, IAM, API keys, Entra). Configure them from the Providers page with the project, region, or resource your organisation requires.
Local runtimes often advertise the training context (for example 128k) while allocating a much smaller window from VRAM. Ollama may truncate oversized prompts silently from the front, which drops system instructions. Configure num_ctx / load context so the declared window matches what you serve, or the agent looks "dumb" when it was truncated.
Self-hosted keeps prompts and tool output off third-party networks. Smaller local models are weaker at long tool-calling loops; a common pattern is local for sensitive material and a hosted model for hard reasoning on sanitised questions.
Catalogue and discovery
Providers and model metadata come from models.dev. A snapshot ships with the extension so the picker works offline; it refreshes in the background when the network is available.
The catalogue is the general offer. Your account may differ. Studio asks each configured provider what it actually serves and reconciles that with the catalogue. What you see in the picker is what your key or session can call.
When a provider retires or renames a model, discovery follows. A missing model usually means it is gone upstream.
Effort and fast mode
Settings → Model. The same knobs are on the model in the composer picker.
Models with a thinking mode can be set to off, low, medium, or high. Fast mode, where the provider offers a faster tier, spends more of that plan's quota for a faster reply. Each is remembered per model, against a default on the Model page. The page lists the models that override the default, and each override can be reset. A card is omitted when no configured model offers that knob.
Reasoning tokens are billed and count against the context window. High effort on a long autonomous run is an easy way to a surprising invoice. Fast mode is on the same bill: the picker marks a model that has it on.
Context
Every turn has to fit in the model's context window. The context gauge shows how full it is.
Exegol System is Studio's built-in layer: workspace layout, approvals, follow-up behaviour, false-positive handling, where tools and wordlists live, aliases, and the rest of the cockpit contract. It keeps the model useful and controlled inside Exegol. It is not an editable system prompt: opening it would put approvals and tool routing at risk. Exegol keeps trimming it; expect incremental gains, not an empty shell you maintain yourself.
Everything else is yours and spends the same window: your messages, personas, the harness (skills, rules, agents, hooks, briefs), MCPs you enable, and attachments or mentions from the composer. A light persona and a lean harness leave more room; heavy always-on rules, large briefs, and many MCP schemas spend it faster.
When the gauge climbs, prefer levers on your side, or a larger window, over trying to gut Exegol System: trim persona / harness / MCP weight, pick a model with a larger window (catalogue), or compact (auto-compaction or /compact; detail is lost).
Writable system prompts are a common request and a common way to break the product. Put methodology in the harness and tone in personas instead.