Local-first Codex preflight

Find unnecessary prompt text before sending, without changing your intent.

Use fewer prompt tokens when the saving is clear and safe.

UseTokenSaver checks a prompt before a Codex turn and stays silent unless savings are high-confidence and material. It never silently rewrites your prompt or changes native Codex settings.

Try the offline preview Download the CLI/SEA package
Offline / CLI verified Local-first · fail-open Desktop host integration unverified

What is verified

The frozen Windows SEA artifact and local analysis path have been checked for deterministic decisions, privacy boundaries, bounded input, fail-open behavior, and supported GPT-5.6 model metadata.

  • GPT-5.6 Sol, Terra, and Luna only
  • No raw prompt retention by default
  • Unknown or unsafe cases silently allow
  • Self-contained Node 24 Windows executable

What is not claimed

Desktop UserPromptSubmit host invocation and output acceptance are not verified. The current Desktop surface exposed configuration identity, but no authoritative process trace for the permitted synthetic capture.

Desktop limitation: do not treat this page as proof of reliable Desktop pre-send interception.
Compute-profile limitation: the current release does not provide a verified recommendation for the most economical model, reasoning effort, and speed combination. GPT-5.6 model metadata is available, but authoritative Codex-credit/profile ratios are not publicly verifiable. The disclosed recommendations concern safe prompt analysis and estimated input savings; native Codex settings are never changed silently.
Open request for guidance

Help us resolve the two remaining Codex evidence gaps

UseTokenSaver’s offline/CLI behavior is verified. Two evidence gaps remain: reliable Codex Desktop Hook invocation evidence, and authoritative Codex-credit/profile evidence for recommending the most economical model, reasoning effort, and speed combination.

If you know the supported Desktop Hook contract, trust/refresh semantics, or an official diagnostic path, we’d welcome guidance. Please don’t share private prompts, sessions, or credentials.
If you know the supported Codex-credit/profile documentation or public evidence needed to compare model, reasoning effort, and speed, we'd welcome guidance. Please don't share private prompts, sessions, credentials, or account data.

This is a request for public guidance, not a claim that Desktop pre-send enforcement or full compute-profile optimization is complete.

Offline decision preview

This browser preview replays disclosed synthetic decision fixtures. It runs locally in your browser, sends nothing to a model, and is not a replacement for the Windows CLI Hook.

Choose a synthetic case to see the disclosed decision shape.

Download and run the verified CLI Hook

Download CLI/SEA package (ZIP)

Read installation instructions · Verify SHA-256 · View synthetic input fixture

The package contains the self-contained Windows executable, license notice, hash file, and synthetic fixture. Verify the digest before installing.

In a supported local Codex Hook environment, the installed command is:

%PLUGIN_ROOT%\dist\usetokensaver-hook.exe

The Hook reads a bounded JSON envelope from stdin and emits only a validated decision when it can prove a material, safe recommendation. Standalone invocations without the required host/session context fail open with empty output and exit code 0.

Repository source, certified build details, and the full test evidence remain private; this page publishes only the product boundary and synthetic behavior.