Rehearse for $0

Documentation

Models and providers

docs/PROVIDERS.md ↗

yaaf runs on any of 23 model providers, including a model on your own machine. Each seat on the org chart can run on a different one. Every run is metered against one budget per project, whoever it paid.

How it fits together

Three ideas:

  1. A credential is a key for one provider, stored encrypted. You add it once, under Settings → Providers.
  2. A project pays with one credential, and can add more for other providers.
  3. A role names a class of model, not a model. @frontier, @balanced or @fast. Each provider says which of its models each class means. Change that once and every seat follows.
Class Meant for
@frontier The most capable model the provider sells
@balanced The one you’d run all day
@fast Cheap and quick: triage, formatting, lookups

A seat can also name an exact model (for example claude-sonnet-5-5), or a model the catalog doesn’t list as provider/model, like openrouter/qwen/qwen3-coder or ollama/llama3.1:8b.

Providers

Provider Key Searches the web itself Classes set for you
Anthropic API key, or subscription token Yes Yes
OpenAI API key, or ChatGPT login Yes Yes
Google (Gemini) API key, or Google login Yes Yes
Google Vertex AI Service account key + project and region Yes Picked from your key
Azure OpenAI API key + resource name Yes No
AWS Bedrock API key + region No No
DeepSeek API key No Yes
Mistral API key No Yes
xAI (Grok), Moonshot (Kimi), Z.ai (GLM), Qwen, Cohere API key No Picked from your key
OpenRouter, Groq, Together, Fireworks, Cerebras, NVIDIA NIM, Hugging Face API key No No
Ollama, LM Studio Address (key optional) No No
Any OpenAI-compatible server Address + key No No

yaaf providers prints this list from the install itself, with what each needs.

“Classes set for you.” Yes: yaaf ships a lineup for it. “Picked from your key”: yaaf reads the models your key can reach and picks by price and age. No: name the model behind each class yourself, in Settings → Models or with yaaf set-lineup PROVIDER CLASS MODEL. A seat whose class has no model is refused before it starts, with a message that says so.

Add a key

Dashboard: Settings → Providers → Add. Then press Check now on it to test it with one real call.

CLI:

yaaf create-llm-credential my-anthropic --provider anthropic
yaaf create-llm-credential local-llama --provider ollama --base-url http://localhost:11434/v1

The value is prompted for and never echoed. Then prove it works:

yaaf smoke <credential-id>            # one tiny real call
yaaf check-provider <credential-id>   # can these models do yaaf's work? Shows the cost first

Give a project its keys

A project pays with one credential, set when you create it or later in the project’s settings. To run some seats on another provider, add a second:

yaaf pay-provider my-project openai <credential-id>

Every run still counts against the project’s one budget.

Subscriptions

Anthropic, OpenAI and Google also take a subscription login instead of an API key. Runs then use your plan’s allowance instead of dollars, and the dashboard shows what is left on the plan.

Provider What you paste Seats run on
Anthropic Subscription token Claude Agent SDK
OpenAI ChatGPT login (auth.json) Codex CLI
Google Google login (oauth_creds.json) Gemini CLI

Check your provider’s terms before running many agents on a personal plan. Anthropic’s consumer terms do not permit it: use subscription mode only for small, single-user tinkering.

Harnesses

The harness is the program that drives the model inside a run. yaaf picks one from the credential; you can override it per role or per seat.

Harness Providers
Claude Agent SDK, Claude Code CLI Anthropic
Codex CLI OpenAI
Gemini CLI Google
OpenCode CLI OpenAI and most OpenAI-compatible providers
yaaf’s own loop Every provider, on its own API

The rules are the same on every harness. A harness decides how a model is called, never what it may do.

Most roles browse. Anthropic, OpenAI, Google, Vertex and Azure search on their own. For every other provider, set one search service for the whole install:

yaaf set-web-search brave           # key prompted
yaaf set-web-search searxng --base-url http://searxng:8080   # self-hosted, no key

Choices: Brave, Tavily, Exa, Serper, Perplexity, or a self-hosted SearXNG. Each search is billed into the run at the price you set (--price). Without a search service, a role that needs the web is refused on a provider that can’t search.

Prices

yaaf meters every run in dollars, so it needs each model’s price.

  • Automatic. Prices come from the LiteLLM price feed, refreshed daily while upkeep is on (the default). A price you set by hand is never overwritten.
  • By hand. yaaf set-model-price MODEL INPUT_USD OUTPUT_USD, per million tokens. Use it for a model the feed doesn’t know.
  • Local models cost nothing: price them at zero, like yaaf set-model-price ollama/llama3.1:8b 0 0. yaaf setup does this for you.
  • No network to GitHub? Turn upkeep off (yaaf upkeep --disable) and run yaaf import-prices --snapshot --apply to load the price list that ships with yaaf.

A model with no price can’t be hired, because its runs couldn’t be counted against the budget.

Limits

yaaf set-provider-ceiling anthropic 4    # at most 4 runs on Anthropic at once, install-wide

Use it to stay under a provider’s rate limit when several projects share one key.

Local models

Point the credential at the server’s address.

  • Native install: http://localhost:11434/v1 (Ollama) or http://localhost:1234/v1 (LM Studio).
  • Docker install: runs happen in containers, where localhost is the container itself. Put the model server on the sandbox network (YAAF_SANDBOX_NETWORK, default yaaf_sandbox) and use its name, like http://ollama:11434/v1.

Local models cost nothing, but small ones often fail at tool use. Run yaaf check-provider before you give one real work.

The Founder’s key

The Founder (the assistant that drafts a project with you) uses its own key:

yaaf set-platform-key --from-credential <credential-id>

Network

Both the server and the agent sandboxes need outbound HTTPS to every provider you use. Vertex AI also needs oauth2.googleapis.com from the server, which mints its tokens there; the service account key never enters a sandbox.

Useful commands

Command What it does
yaaf providers Every provider, and what it needs
yaaf models What each class means on each provider
yaaf discover-models CRED What models a key can reach, with prices
yaaf repick-lineup Move a picked class onto a newer model
yaaf set-default-provider The provider used before a project pays one
yaaf spend --by provider What each provider cost