Show HN: AI-whisper – two real coding agent CLIs, one implements, one reviews AI-whisper, a new open-source tool, pairs two coding agent CLIs—such as Claude, Codex, or ezio—in a terminal-native workflow where one agent implements code while the other reviews it, autonomously iterating until a deliverable is approved. The tool targets engineers who already use coding agents and want structured, multi-agent review loops without manual oversight. ai-whisper pairs two coding agents — mount any two of Claude, Codex, and ezio — into a terminal-native pair that hand work back and forth under a single baton, so one agent implements while the other reviews, and a structured workflow drives the loop to a finished, reviewed deliverable without a human babysitting every round. Mount each agent in its own terminal. Each mount claims the current shell, launches the real provider CLI, and binds it to the collab: terminal 1 whisper collab mount claude terminal 2 whisper collab mount codex Then, from inside either agent's session, kick off a structured workflow against a spec — just ask in plain language: Run spec-driven-development using docs/spec.md From there ai-whisper runs the workflow autonomously: Implementer / reviewer assignment — the agent you trigger the workflow from becomes the implementer and the other agent becomes the reviewer; pass --implementer / --reviewer to choose explicitly. Started outside a mounted session with no flags, it falls back to a default pairing and warns. The baton passes between them; only one owns the turn at a time. Autonomous execution — the implementer does each step in its real session and hands the result back. An LLM evaluator judges whether the deliverable meets the request. Review loops — when work isn't good enough yet, the reviewer's findings are composed into a follow-up handoff and the implementer iterates. The loop repeats until the work is approved or the round budget is exhausted. Resumability — workflow and chain state is durable. If the broker restarts or you stop for the day, you recover and reconnect rather than starting over. Deliverables — you get committed code plus a review trail per-step verdicts, round counts , inspectable at any time with whisper collab dashboard . A real spec-driven-development run: Claude left and Codex middle work in their own mounted sessions while the dashboard right tracks the baton handoffs and per-phase verdicts ~20s . Click the still to watch it play on the project page. ai-whisper is for engineers who already lean on coding agents and want more structure around them: - you already use coding agents heavily and want two of them to check each other. - you work terminal-first and want the agents to live in real terminal sessions, not a web UI. - you want multi-agent review — a second model gating the first model's output. - you run long, structured workflows spec → plan → implement → review rather than one-off prompts. It is not for: - one-shot "vibe coding" where you just want a quick answer. - invisible background automation you never watch. - people new to coding agents looking for a guided, hand-holding experience. You pair any two of three agents — claude , codex , and ezio . ai-whisper drives the real Claude and Codex CLIs, so install and authenticate whichever of those two you plan to mount first; ezio is protocol-native and ships with ai-whisper, so it needs no separate CLI. — the Claude Code CLI https://claude.com/claude-code claude command, signed in.— the Codex CLI https://github.com/openai/codex codex command, signed in. ezio optional — bundled with ai-whisper; mount it with whisper collab mount ezio , no separate install. Node.js 22+ . An LLM evaluator with credentials — workflows are gated by it and refuse to start without it. See Evaluator configuration /ai-creed/ai-whisper/blob/master/docs/evaluator-configuration.md . tmux optional — only for whisper collab start , which auto-arranges both agents into panes. The mount flow below does not need it. Platform support:ai-whisper is terminal-native and Unix-oriented — it drives interactive PTY sessions, so it runs onmacOS and Linux. It isnot supported natively on Windows: whisper collab mount / reconnect require a Unix tty-backed shell and will exit with an error pointing here. On Windows, run ai-whisper inside— install Node, your agent CLI, and ai-whisper inside the WSL2 distro and run the commands there, where everything works as-is. WSL2 ai-whisper launches each agent in full-autonomy mode so the relay can drive it unattended — claude --dangerously-skip-permissions and codex --dangerously-bypass-approvals-and-sandbox . Inside the mounted workspace the agents read, write, and run commands without prompting. Point it at code you're willing to let two agents change autonomously, watch the run on the dashboard, and remember you are the final gatekeeper — review the result before you ship it. The deeper rationale is in Concepts /ai-creed/ai-whisper/blob/master/docs/concepts.md . Install from npm: npm install -g ai-whisper Or from a repo checkout: pnpm install pnpm build Install the bundled agent skills once they let the agents verify, kick off, and report on workflows . This also installs ai-whisper-code-review , the skill workflow code-review handoffs use to evaluate agent-written code, and ai-whisper-plan-execution , the skill plan-execution handoffs use to structure how the implementer executes an approved plan: whisper skill install Workflows require an LLM evaluator with credentials — set this up before running one. See Evaluator configuration /ai-creed/ai-whisper/blob/master/docs/evaluator-configuration.md . Then mount both agents and run a workflow: terminal 1 whisper collab mount claude terminal 2 whisper collab mount codex The first mount creates the collab and starts the broker daemon for the workspace; the second binds the other agent. From either session, start a workflow against a spec or goal file — spec-driven-development for a spec, ralph-loop for an open-ended goal, plus complex-bug-fixing and deliberation see Workflows /ai-creed/ai-whisper/blob/master/docs/workflows.md . Watch it run with: whisper collab dashboard whisper collab dashboard — live wall of recently-active collabs + per-run inspector. Add --all to show every workflow run no per-collab masking ; combine with --window all for the full run ledger. Running from a repo checkout instead of a packaged install? Build first pnpm build and invoke the CLI as node packages/cli/dist/bin/whisper.js ... wherever these examples say whisper ... . A run that stops short usually escalates — it does not crash. When the evaluator can't resolve a phase the round budget is spent, an agent reports it's blocked, or confidence stays too low , the loop halts and turn ownership returns to you. That's a designed exit, not a failure: run state is durable, so you read the dashboard, fix the spec or unblock the agent, and whisper workflow resume