Model settings

Connect a service

Open Model settings at the bottom of the sidebar. Choose a provider, enter a model ID available to your account and an API key if required, then click Connect model. A project is optional for conversation and attachment analysis. Open a workspace for development tools.

Hikari includes presets for OpenAI, Azure OpenAI / Foundry, Anthropic, Google Gemini, DeepSeek, Qwen, Kimi, GLM, Doubao, MiniMax, SiliconFlow, Groq, Grok, Mistral, OpenRouter, Ollama and LM Studio. OpenAI compatible accepts other Chat Completions services. Provider selection supplies a starting address; you can change it for your region or private gateway.

Service	Base URL
OpenAI	https://api.openai.com/v1
Azure OpenAI / Foundry	Your resource URL ending in /openai/v1
Anthropic	https://api.anthropic.com/v1
Google Gemini	https://generativelanguage.googleapis.com
DeepSeek	https://api.deepseek.com/v1
Qwen	https://dashscope.aliyuncs.com/compatible-mode/v1
Kimi	https://api.moonshot.ai/v1
GLM	https://open.bigmodel.cn/api/paas/v4
Doubao	https://ark.cn-beijing.volces.com/api/v3
MiniMax	https://api.minimax.io/v1
SiliconFlow	https://api.siliconflow.cn/v1
Groq	https://api.groq.com/openai/v1
Grok	https://api.x.ai/v1
Mistral	https://api.mistral.ai/v1
OpenRouter	https://openrouter.ai/api/v1
Ollama	http://127.0.0.1:11434/v1
LM Studio	http://127.0.0.1:1234/v1

Use the endpoint supplied by your account. Qwen supports workspace and region-specific endpoints; replace the preset when required. Kimi keys differ between regional platforms. Doubao may require a deployment ID. Azure uses the v1 API with an API key and a deployment name in Model; older versioned endpoints and Entra sign-in are not supported by this preset. Anthropic and Gemini use their native protocols; other presets use Chat Completions. Services requiring AWS request signing, Vertex OAuth or a different API need a compatible gateway.

Choose Load available models to fetch the service's model list, then select an ID from the Model field suggestions. A failed or unavailable list does not prevent manual entry. Hikari does not assume that every listed model supports tools or images.

Set model capabilities

• Enable This model supports tool calls for project tasks. Disable it for conversation-only models; Hikari will not send tool schemas to them.
• Enable This model accepts images only for a model that supports image input. Text-only models can use local OCR text when available.
• Turn off Stream responses if your endpoint only supports ordinary responses.
• Set Context window (tokens) and Maximum output tokens to the limits of the selected model. Output space is reserved from the context window.
• Leave Reasoning effort at Provider default, or select a supported value. Unsupported settings may be rejected by the provider.

Test connection uses a small, billable request and checks whether a tool call is returned. It cannot prove the model's overall development quality. A service error stays an error; real tasks never switch to sample answers.

Save and switch profiles

Use New model profile, optionally give it a name, and connect. The Saved model profiles list lets you select and activate a saved configuration with Connect model. Changes apply to new tasks. A resumed task retains its original profile; restore that profile's model and endpoint if they changed, or use Retry to start with another model.

Credentials use OS secure storage when available. Otherwise they stay in memory for the current session. Keys are not returned to the settings page, exported with conversations, or synced. Removing a profile requires its unfinished tasks to be cancelled or completed first.

Optional semantic search

Open Semantic retrieval, enable it and enter an Embedding model, Embedding endpoint and separate Embedding API key. The endpoint must provide OpenAI-compatible embeddings; a localhost service may allow an empty key. Chat and embedding models may come from different services.

Search combines keyword and semantic matches in attached documents and saved project memory. Results include their source labels. Embedding searches require approval before text is sent to the endpoint. Vectors are cached locally and refreshed when source text, endpoint or model changes. Keep an endpoint's vector dimensions stable; use a different model ID when its dimensions change.

This option sends the searched document collection and project memories to your embedding service, not your entire code repository. Collections are limited to 2,000 chunks and 3 million characters; split large collections or disable embeddings for keyword search. Embedding charges are separate from the chat token/cost budget. Keyword retrieval works locally when this option is off; an enabled service failure is reported explicitly.

Offline use

Choose Use offline demo to try sample Q&A without a model or workspace. Existing demo conversations remain simulated when continued or retried. Connect Ollama or LM Studio to use a real local model offline, subject to its tool and image capabilities.