Documentation / The agent

Models and providers

Supported providers in Platform mode, local models, extended reasoning, context, caching, usage and budgets.

In Platform mode, Yuli's engine reasons with the provider and model you choose, using your own API key. This page explains how to configure it and how to keep usage under control.

If you'd rather use your Claude Code, Kiro, Codex or Antigravity subscription, you don't need any of this. See Connect an AI.

Providers

Provider What for
Anthropic The Claude API.
OpenAI GPT models.
Gemini Google AI Studio.
xAI Grok.
DeepSeek The DeepSeek API.
Local An OpenAI-compatible server on your machine or your network: Ollama, LM Studio, vLLM or llama.cpp.
AWS Bedrock Claude, Llama, Mistral, Nova… with your AWS account.
Vertex AI Gemini and Claude with your Google Cloud project.

Configure a provider

Open AI providers in any of these ways:

  • Chat header → More actions → Model settings.
  • Global options (Ctrl+Alt+S) → AI and models → Open model settings.

Then:

  1. Choose the provider's card.
  2. Enter the API key and the Model. For Local, also enter the Server URL.
  3. Click Test connection (optional).
  4. Turn on Use as active provider.
  5. Click Save settings.

Credentials are stored in the system keychain and are never shown in full. To change another field without retyping the key, leave the key field empty: the current one is kept.

AWS Bedrock and Vertex AI

Provider Ways to authenticate
AWS Bedrock Profile / SSO, IAM keys, IAM role or API key.
Vertex AI ADC, Service account, gcloud CLI or API key.

Fast model

Optionally, set a Fast model for lightweight tasks such as summaries and titles.

Local models

With Local, the AI's context never leaves your network. Start your server (Ollama, LM Studio, vLLM or llama.cpp with its OpenAI-compatible API), enter its URL and the model name, and test the connection.

A local model needs strong tool-calling ability to operate the IDE. Try a small task first.

Extended reasoning

In the same panel, Extended reasoning offers Off, Low, Medium or High.

  • It applies only to the agent's main work, not to summaries or titles.
  • Only on models that support it: reasoning budget on Claude (API, Vertex and Bedrock) and Gemini 2.5, and reasoning effort on OpenAI's reasoning models.

Context

Yuli manages the context window for you:

  • It knows each model's context size.
  • The active file, the working diff, the rules and the context bundle are capped. Whatever doesn't fit is named so the agent can read it if it needs to.
  • Long conversations are compacted with incremental summaries.
  • Only the tools that apply to the project are sent (for example, Web Studio's only if there's a front end).

Prompt caching

Automatic. The system prompt is split into a stable part and a per-turn part, and only the stable part is cached. Each message's context travels with that message, so the rest of the conversation keeps being read from the cache. With Claude models, each round within a task reuses everything before it.

Savings

  • An identical call with no changes in between isn't repeated.
  • Independent reads are resolved in parallel.
  • Anything that requires approval runs as soon as you approve it, with no round trips to the model.
  • The provider's retry-after is honored during an outage.

Usage and cost

Where What you see
Status bar Tokens for the session and the current task, and requests. Click for the details.
This session's usage Input, Output and Total. These are the figures each provider reports; they're approximate.
Agent → AI Requests (recent)… Each request with tokens, cost, latency and errors.

In Subscription mode, the bar shows plan usage (5-hour window and week) instead of the token count.

Budgets

Where Setting
Verification button in the chat header → Auto-verify Max iterations and Budget per task (USD).
More actions → Agent team Budget per task (US$ 8) and Budget per mission (US$ 60).
Global options → AI and models → Background tasks See below.

Background tasks

Setting Default
Concurrent tasks 2 (from 1 to 8)
Tokens per task 2,000,000
Cost per task (US$) 5
Fix attempts after tests 2

Save with Save limits. When a task hits its cap, it pauses with a usage bar and you can Raise limit and resume.

AI autocomplete

Inline autocomplete from the local engine is always available. AI autocomplete ships off and is enabled by choice, with the provider and model you prefer.