Models and providers

Abhed does not ship a model. It is model-agnostic by construction: the harness is the same whichever endpoint you point it at, and a better model makes the same harness better.

abhed providers lists what your binary supports — which is what registered itself at build time, so a build that drops the cloud adapters for an air-gapped install reports honestly.

The providers

Three wire formats cover all of them. The OpenAI-shaped providers share one adapter and differ only in default URL and which sampler knobs the server honours; Anthropic and Gemini are genuinely different shapes.

Sampling

"params": {
  "temperature": 0.2,
  "top_p": 0.9,
  "top_k": 40,
  "seed": 42
}

An omitted parameter is not zero. Leaving temperature out keeps the model's own default; setting it to 0 asks for greedy decoding. Collapsing those would make "don't touch it" impossible to say.

A parameter the provider cannot honour is refused at startup, naming the knob and the provider. min_p sent to a hosted API is ignored silently and the evidence is nowhere in the output — the answers are simply drawn from a distribution nobody chose. Failing the config is the smaller harm.

Precedence is request → provider params → model default.

Authentication

An API key, from config or the environment:

{ "type": "anthropic", "model": "claude-opus-5", "api_key_env": "ANTHROPIC_API_KEY" }

Subscriptions do not work, and this is not a Abhed limitation

A Claude Pro or Max token — the kind claude setup-token prints — is restricted to Claude Code. Anthropic accepts the credential and then refuses the request unless the system prompt is Claude Code's own. Measured directly: same token, same model, same second, the only difference being the first system block.

The refusal arrives as 429 rate_limit_error, which is misleading — it is a policy decision, not a limit that clears. Abhed recognises the shape (a 429 carrying none of the headers a real rate limit carries), reports it as what it is, and does not retry.

Abhed reads CLAUDE_CODE_OAUTH_TOKEN and oauth_token because the mechanism is correct and the restriction may not be permanent. Today it is useful only for models outside the check.

Working around it means sending Claude Code's identity string from a product that is not Claude Code. That circumvents an access control, misrepresents the product, and breaks the moment the check changes — so Abhed does not do it, and neither should anything built on it.

Some models outside the check still answer. That is not a reason to rely on it: third-party usage of a subscription is billed as extra usage rather than drawn from the plan, so a path that looks free is metered somewhere the Console does not show — a subscription token is not a Console credential, and spending under it appears only in claude.ai settings.

Use an API key. From platform.claude.com, billed per token, no gate, and visible where you would look for it:

export ANTHROPIC_API_KEY=sk-ant-api03-...

Adding a provider without a rebuild

"custom_providers": [
  {
    "name": "internal-vllm",
    "api": "openai",
    "base_url": "https://llm.internal.example/v1",
    "sampling": ["temperature", "top_p", "top_k", "seed"]
  }
]

api is named rather than guessed from the URL: a wrong guess produces rejected requests whose errors point nowhere near the cause.

Switching mid-session

/model <name> swaps the provider and keeps the conversation. The next turn re-prefills, because the new provider has never seen this prefix — a real cost, and still cheaper than rebuilding the session by hand.

Choosing one

abhed doctor tells you whether a model can drive the agent at all. abhed-modelcmp compares candidates on tool calling and explanation quality, which are different things and are not measured by the same benchmark.

Decode speed tracks active parameters, not total: a 27B dense model can be far slower than a 30B mixture-of-experts.