Choose a connection¶

The first decision is where provider access and assessment state should live.

Decision guide¶

Local Core¶

Start the Core in a terminal:

stealth-prompt serve

It binds to loopback, prints a one-time pairing code, and never opens a browser. Enter the code in Setup → Local Core. The extension exchanges it for a token bound to the extension origin. Restarting the Core invalidates the token.

The Core discovers available providers and models:

• Fake is deterministic, offline, and intended for the demo and workflow checks.
• Claude CLI uses the installed and authenticated claude executable.
• Codex CLI uses the installed and authenticated codex executable.
• Ollama uses a loopback endpoint and refuses a remote URL.
• OpenAI API reads STEALTH_PROMPT_OPENAI_API_KEY, then OPENAI_API_KEY.

Run stealth-prompt doctor when a CLI provider is missing or model discovery fails. Use stealth-prompt serve --port <port> for a non-default port and enter the same port in the extension. Core also owns the private, revisioned strategy library. It uses .stealth-prompt/strategies.sqlite3 by default; pass --strategy-db <file> to keep it elsewhere. Strategy snapshots contain reusable abstractions, not captured replies or credentials. Before each planning turn, Core filters active strategies by objective (or searches across objectives when Auto is selected), scope, mode, observed surface/capabilities, response pattern, and exhausted moves. It then exposes at most three bounded strategy summaries to the planner. Target and project scope keys are SHA-256 identifiers rather than stored origins.

Core can also learn from a completed report, but only when the operator enabled Offer this run for reviewed learning before that run. Reports shows eligibility and, on request, generates a sanitized preview using the source report's provider and model. Accept, Edit, and Keep target-specific are the only actions that mutate the local library; Reject records only the decision. Reopening an accepted report is idempotent. Reviewed libraries can also be evaluated against a frozen local benchmark, promoted through fixed quality gates, and exchanged as explicitly trusted signed files. See Evaluate and share strategies.

Direct API¶

Choose Direct API, then OpenAI or Anthropic. Enter a restricted project key and select Use key & load models. Chrome requests access only to the selected provider's fixed API origin.

The key is held in Side Panel and service-worker memory only while needed. It is not written to chrome.storage, bindings, reports, or exports. Closing the panel clears it.

Direct API supports Payload only, Assist, Guided, and bounded Auto. It does not provide CLI providers, Core scenarios, deterministic scorers, the Core's HTML artifact, or a mutable private strategy library. It does use the shipped read-only strategy catalogue, with the same bounded selection and report attribution.

Direct reports can be reviewed and exported in the extension, but they are not learning inputs and never mutate a paired Core. Reports shows a quiet link to the Core setup when you want the private, revisioned library.

FrameFuzz is available on both paths. Core can evaluate cases with deterministic scorers and can therefore reach a confirmed framing-gap conclusion. Direct API uses the same local deterministic templates and fresh-context gates, but its conclusion is capped at potential because its evaluator is a model. See FrameFuzz.

Data sharing¶

The selected policy controls whether captured target responses reach the provider:

Claude, Codex, OpenAI, and Anthropic may be external services under your own provider agreement. Ollama is restricted to loopback. Stealth Prompt does not change provider retention policies.