Providers and models

Yachiyo has no built-in model and no hosted account. You bring providers, and you can have as many as you like at once — switching per message is normal usage, not an edge case.

Providers live in ~/.yachiyo/config.toml. API keys are stored there in plain text on your machine and are redacted in all CLI output.

Adding a provider

Settings → Providers → Add, then pick a preset. The preset fills in the API type and base URL; you supply the key.

Preset	API type	Base URL
Anthropic	anthropic	https://api.anthropic.com/v1
OpenAI	openai	https://api.openai.com/v1
OpenAI (Codex OAuth)	openai-codex	https://chatgpt.com/backend-api/codex
Gemini	gemini	https://generativelanguage.googleapis.com/v1beta
Google Vertex AI	vertex	— (uses project + location)
Vercel AI Gateway	vercel-gateway	https://ai-gateway.vercel.sh/v3/ai
OpenRouter	openai	https://openrouter.ai/api/v1
DeepSeek	openai	https://api.deepseek.com/v1
Moonshot	openai	https://api.moonshot.cn/v1
Kimi For Coding	anthropic	https://api.kimi.com/coding/v1
Zhipu GLM	anthropic	https://open.bigmodel.cn/api/anthropic
Z.ai	anthropic	https://api.z.ai/api/anthropic
Minimax	openai	https://api.minimaxi.com/v1
Mistral	openai	https://api.mistral.ai/v1
Aliyun Bailian	openai	https://dashscope.aliyuncs.com/compatible-mode/v1
PackyCode	anthropic	https://www.packyapi.com/v1
Ollama	openai	http://localhost:11434/v1

Anything not on the list works too — pick the preset whose API shape matches (most third-party endpoints are OpenAI- or Anthropic-compatible) and replace the base URL.

Local models

The Ollama preset points at http://localhost:11434/v1 and needs no API key. Any other OpenAI-compatible local server (llama.cpp, LM Studio, vLLM) works the same way — set the base URL and leave the key blank if the server does not check it.

Vertex AI

Vertex does not use an API key. It needs project, location, and a service account (serviceAccountEmail + serviceAccountPrivateKey) instead.

Codex OAuth

The openai-codex type authenticates through a Codex CLI session rather than a key. Point codexSessionPath at your Codex auth.json and Yachiyo reads the account from it.

Choosing which models appear

Fetching a provider’s catalogue returns everything the API knows about, which is usually far more than you want in a picker. So each provider keeps two lists:

• modelList.enabled — shows up in the model picker
• modelList.disabled — known, deliberately hidden

Only enabled models are selectable. Curate the list once and the picker stays short.

yachiyo provider models openai      # everything the API offers
yachiyo provider models             # what you have actually enabled

There is also modelList.imageIncapable — models are assumed to accept images, so list the ones that do not.

The default model

The defaultModel setting decides what new threads start with:

yachiyo provider set-default anthropic --model <model-name>

That promotes the provider to default and sets the active model in one step. Without --model it picks the provider’s first enabled model. Individual threads and messages can override it freely.

The tool model

Separate from the model that talks to you, Yachiyo uses a tool model for background work:

• thread titles
• memory generation and distillation
• context handoff when a thread outgrows its window
• image-to-text descriptions
• group-chat probes and reply rewriting
• the built-in translator

Three modes:

Mode	Behavior
default	Reuse the chat model. Simple, and the most expensive option.
custom	Name a specific provider and model.
disabled	Skip auxiliary generation entirely. Titles stay generic, distillation stops.

This work is frequent and low-stakes, so custom pointed at something small and fast is usually the right setup.

yachiyo config get toolModel
yachiyo config set toolModel.mode '"custom"'

Reasoning effort

For models that expose a reasoning budget, each provider can carry a reasoning block: which effort levels are offered (low, medium, high, xhigh, max), which is the default, and whether off is allowed. Effort is then selectable in the composer per message.

Managing providers from the CLI

yachiyo provider list                       # all providers, keys redacted
yachiyo provider show openai                # one provider in full
yachiyo provider update openai --payload '{"apiKey":"sk-..."}'
yachiyo provider set-default openai --model <model-name>

Full details in the provider CLI reference.

Keys in config.toml

config.toml stores API keys in plain text. Do not commit it and do not put it in a shared folder.

CLI output is safer than the file: every command replaces apiKey values with ***, including yachiyo config get. Reading the file directly is the only way to print a key.