Rehearse for $0

Harnesses and models

Every harness. Every model. One company.

An agent harness drives one model. yaaf is the harness around the harnesses: put each person on the chart on the harness and model that suits the job, and the whole team still answers to one budget and one set of rules.

6

Harnesses

23

Providers

26

Models in the catalog

1

Budget per project

§ Two levels

A harness is transport. yaaf is the rules.

Claude Code, Codex and the rest decide how a model is driven. They never decide what it may do. That lives in yaaf's server, the same for every harness.

The agent harness decides
  • How the model is called, and its loop
  • The tools it offers the model
  • How a long run is compacted to fit
yaaf decides, for every harness
  • Who is hired, and who they report to
  • What may be worked on, and when it's done
  • What it costs, metered live against one cap
  • When it's due, and what's on the record

A harness that can't report what it spent while it's spending it isn't let in. Every one runs inside yaaf's sandbox, on yaaf's meter.

§ Model classes

Roles name a class, not a vendor.

A role asks for frontier, balanced or fast. Each provider says which of its models that means, so the same team runs on whichever provider you pay.

Frontier

The most capable model the provider sells. For the CEO and the hardest seats.

Balanced

The workhorse. Most of a team runs here.

Fast

Cheap and quick, for the seats that read and sort.

Pin a seat to an exact model whenever you want. If a provider keeps failing, a project can name a fallback credential: its class seats move there until the provider recovers, and back. Pinned seats never move.

§ Why

Two agents on one model share its blind spots.

Critics on a different model

Put your researcher on Claude and its critic on GPT. Disagreement gets real.

Not tied to one vendor

The rules never depended on the model. Now the metering doesn't either.

§ Harnesses

What drives the model.

HarnessTypePays withVersion
Claude Agent SDKSDKAnthropic API key · Claude subscriptionlatest
Claude Code CLICLIAnthropic API key · Claude subscriptionlatest
yaaf agent loopAPI loopAnthropic API key · OpenAI API key · Gemini API key · Vertex AI service account · Azure OpenAI API key · AWS Bedrock API key · DeepSeek API key · Mistral API key · xAI (Grok) API key · OpenRouter API key · Groq API key · Together AI API key · Fireworks AI API key · Cerebras API key · Cohere API key · NVIDIA NIM API key · Hugging Face API key · Z.ai (GLM) API key · Moonshot (Kimi) API key · Qwen (Alibaba Model Studio) API key · Ollama (local) · LM Studio (local) · Custom endpoint key—
Codex CLICLIOpenAI API key · ChatGPT plan (Codex login)0.162.1
Gemini CLICLIGemini API key · Google account login (Gemini CLI)0.63.0
OpenCode CLICLIOpenAI API key · DeepSeek API key · Mistral API key · xAI (Grok) API key · Groq API key · Together AI API key · Fireworks AI API key · Cerebras API key · Cohere API key · NVIDIA NIM API key · Hugging Face API key · Z.ai (GLM) API key · Moonshot (Kimi) API key · Qwen (Alibaba Model Studio) API key · Ollama (local) · LM Studio (local) · Custom endpoint key1.19.0

Every person on the org chart runs in one of these. Mix them freely on one team.

§ Providers

Who sells the tokens.

ProviderCredentialsThrough
AnthropicAnthropic API key · Claude subscriptionClaude Agent SDK · Claude Code CLI · yaaf agent loop
OpenAIOpenAI API key · ChatGPT plan (Codex login)yaaf agent loop · Codex CLI · OpenCode CLI
Google (Gemini)Gemini API key · Google account login (Gemini CLI)yaaf agent loop · Gemini CLI
Google Vertex AIVertex AI service accountyaaf agent loop
Azure OpenAIAzure OpenAI API keyyaaf agent loop
AWS BedrockAWS Bedrock API keyyaaf agent loop
DeepSeekDeepSeek API keyyaaf agent loop · OpenCode CLI
MistralMistral API keyyaaf agent loop · OpenCode CLI
xAI (Grok)xAI (Grok) API keyyaaf agent loop · OpenCode CLI
OpenRouterOpenRouter API keyyaaf agent loop
GroqGroq API keyyaaf agent loop · OpenCode CLI
Together AITogether AI API keyyaaf agent loop · OpenCode CLI
Fireworks AIFireworks AI API keyyaaf agent loop · OpenCode CLI
CerebrasCerebras API keyyaaf agent loop · OpenCode CLI
CohereCohere API keyyaaf agent loop · OpenCode CLI
NVIDIA NIMNVIDIA NIM API keyyaaf agent loop · OpenCode CLI
Hugging FaceHugging Face API keyyaaf agent loop · OpenCode CLI
Z.ai (GLM)Z.ai (GLM) API keyyaaf agent loop · OpenCode CLI
Moonshot (Kimi)Moonshot (Kimi) API keyyaaf agent loop · OpenCode CLI
Qwen (Alibaba Model Studio)Qwen (Alibaba Model Studio) API keyyaaf agent loop · OpenCode CLI
OllamaOllama (local)yaaf agent loop · OpenCode CLI
LM StudioLM Studio (local)yaaf agent loop · OpenCode CLI
Custom endpoint (OpenAI-compatible)Custom endpoint keyyaaf agent loop · OpenCode CLI

6 models ship with prices, all from Anthropic. Price any other model in Settings. Zero is fine for your own server.

§ 8 rules

Checked when a seat is hired, and again before every run.

R1

The seat's harness must accept the credential.

R2

The harness must honour the role's tool permissions.

R3

If yaaf prices the run, the model needs a price.

R4

One project, one unit: dollars or tokens, never both.

R5

The project needs a credential for the seat's provider.

R6

The model must support the effort level asked for.

R7

The CEO's harness must be able to hold a conversation.

R8

A seat that names a class needs a model for that class on its provider.

§ Status

Built and tested. Not yet proven live.

Done

  • Every harness passes its contract tests, offline
  • Every part of yaaf runs on any provider: the seats, the CEO chat, the Founder and the goal grader
  • Roles name a class, and prices import from a feed
  • Mixed providers on one budget is a test, not a promise

Not yet

  • A live seat on each provider, which only a deployment with real keys can prove. yaaf smoke checks one on yours.
  • A Cursor CLI driver

Harnesses, providers and keys →