Settings

Providers

Models are fetched live from the provider when you connect, with a hardcoded fallback list if the API is unavailable.

AWS Bedrock accepts Bedrock API keys, IAM access keys, or ambient AWS credentials — environment variables, ~/.aws/credentials, or IAM roles.

Prompt caching is applied per provider, automatically, and there is nothing to set for it here; SPEC4_PROMPT_CACHING=0 in the environment turns it off, for measurement. Caching

Where keys live

The key you enter is held in your browser and sent to one place: the provider it belongs to. It is never sent to spec4.ai or anywhere else but your provider.

Two browser stores are involved. The session store (sessionStorage) holds the key for the current tab. If you tick Remember provider and keys in this browser, the preferences store (localStorage) keeps it across sessions and upgrades; that is the opt-in. Clear saved credentials empties the preferences store. The server process holds no credential.

Default model and effort

The model step sets the project's default: one provider, one model, and one reasoning effort. Every agent uses the default unless you choose otherwise for that agent (next section).

Effort is sent to the provider as LiteLLM's reasoning_effort. The values offered are default, low, medium, and high, plus max on Anthropic models and xhigh on OpenAI models. Whether a model accepts the parameter at all is checked with a network-free capability probe; which levels it accepts is not, so:

• default sends medium on a model that accepts the parameter, and nothing on one that doesn't. On models that think whenever nothing is sent, sending nothing meant thinking at the API's own depth — minutes of silence on a large prompt, with the thinking left out of the stream. A level bounds the thinking and makes its summaries stream.
• A model that doesn't accept the parameter ignores it; the call goes through.
• A model that accepts the parameter but rejects the level is retried once with no effort, and the call is recorded in usage.json as default (fallback from <level>).

While a model reasons, the screen counts it, in every agent. In CodeScanner, Brainstormer, Agentifier, StackAdvisor, Phaser, and Deployer, the status line under the composer reads Thinking — N chars as the reasoning streams, until the first character of the reply arrives; Agentifier's count includes the reasoning of its sub-agents. Designer's counter shows the same while it waits for the mock's first output, and also reports a pause of more than 15 seconds mid-draw with how long it has lasted, so a long wait reads as a pause rather than a hang.

Spec4 never sends a provider-specific thinking or budget parameter. reasoning_effort is the only mechanism.

Per-agent model and effort

Each agent can run on a different provider, model, and effort from the default, chosen in three places:

• The gate. Before an agent starts, a one-line panel names the agent and what it is about to run on — Model for Phaser: claude-sonnet-5 · default — with three choices: use the default, pick a model, or keep the override carried forward from the agent's last run.
• The chip. Under the composer while an agent is running, Model: <name> · Change reopens the choice mid-agent.
• The retry. When a provider returns an error, the retry panel offers a different provider, model, and effort. What you pick becomes the agent's selection for the rest of the round.

An override holds its own key and never changes the default. Sub-agents inherit their parent's choice — Brainstormer's feature speccer runs on Brainstormer's model, Phaser's seam check on Phaser's. The status bar, the chip, and each agent's row on the project page show the selection as <model> · <effort>, or the model alone when the effort is default, and usage.json records what every call actually ran on.

The developer picks, every time. Nothing in Spec4 routes a call to a model on its own.

Search

Web search is optional and off unless you give it a key. With one, every agent can search for the standards, protocols, and SDKs it names and embed the canonical documentation link rather than recall one.

Both are reached over MCP; the key is sent only to the provider it belongs to. The key is stored like the provider key, in the same two browser stores under the same opt-in.

What the published rounds ran on

Spec4's three rounds on itself, from each round's usage.json:

• Default: claude-sonnet-5 on Anthropic, in all three rounds.
• Phaser: claude-opus-5, in all three rounds.
• Designer: claude-fable-5-1 in v0 and v1; claude-opus-5 in v2.
• Every other agent that ran — CodeScanner, Brainstormer, StackAdvisor — on the default.

SnarkCheck's one round, from its usage.json:

• Brainstormer and Agentifier: claude-opus-5 at default effort.
• Designer, StackAdvisor, Phaser, and Deployer: claude-sonnet-5 at medium.

That is practice, not a recommendation. Phaser got the stronger model because it holds every upstream artifact in context while it drafts and was the first or second largest line in every round; Designer got one because its single call is the whole mock. Spec4's three rounds predate per-agent effort, so every call ran at the provider's default; SnarkCheck's records it on every call. SnarkCheck's allocation is the reverse of Spec4's — the stronger model at the front, where the product decisions are made, and the default at medium on everything after, Phaser included, for $0.43. Neither is a recommendation. Per-agent figures are on the Artifacts page.