Bouncer

Appearance

Changelog ​

Bouncer is pre-release (0.1.0, installed from source). The full history is in the commit log.

0.1.0 · October 2026 ​

Saving tokens ​

  • Eco mode: once a session passes its budget (200k by default), a new task starts in a fresh session with a handover on your clipboard. Follow-ups are never interrupted.
  • Opus guard from the first prompt, in any language, and only while the context is under 50k (switching in a bigger session costs more than it saves). Paste to keep Opus, or bouncer opus warn to never be stopped.
  • Model and effort advice after each answer: TRIVIAL → Haiku, STANDARD → Sonnet, COMPLEX → Opus, with the command to type.
  • Prompt checks in any language: vague asks, unscoped repo-wide reviews, oversized pastes and log dumps.
  • Lean mode: the optional AI only runs when a decision depends on it (~0.4 s per prompt instead of ~2.9 s).

Measuring honestly ​

  • A/B measurement (bouncer measure on), comparing real token usage only at the same context size.
  • Money estimates removed: "tokens saved" is labelled as an estimate.

Safety ​

  • Leaked secrets are blocked before any model sees them, even with !.
  • AI file hints are checked against git and dropped when unrelated or unsure.
  • Fail-open: any error lets the prompt through.
  • Bouncer's own handover is never blocked when you paste it back.

Where it runs ​

  • Claude Code CLI, desktop Code tab, VS Code and JetBrains.
  • OpenAI Codex (bouncer init --target codex).
  • Free AI presets: Cloudflare Workers AI (Mistral Small 3.1), Groq, Gemini, Ollama.

Rescue and reports ​

  • Rate-limit rescue: a git-backed handover with your verbatim prompts.
  • After-answer report with context bar, status and next steps; always-visible status line; branded block messages with the next step in capitals.
Was this page helpful?