Local-first · No signup · Built in 2 hours
Your agents finished 12 tasks. You reviewed 0.
Agent Inbox pulls the background output from Claude Code, Codex and Cursor into one reviewable inbox. Paste a log, decide accept or ignore in 30 seconds, export what you kept.
No spam. One email when the Pro build ships. Unsubscribe in one click.
Why Agent Review Matters Right Now
The Show HN thread for Pizza Bot — an inbox for AI agents that work in the background — hit the front page with 56 points and 33 comments in a single day. That thread is the signal: developers are no longer running one assistant in one tab. They are running three, four, sometimes six coding agents in parallel, each one editing files, opening pull requests and writing logs while nobody is watching. The work is happening. The review is not.
The tools that exist today solve the wrong half of the problem. Orchestrators like Amika, Agenttik and Legion focus on launching and coordinating agents, and Pizza Bot itself is a delivery protocol — it defines how an agent pushes a result into a queue. None of them answer the question a developer actually asks at the end of the day: which of these twelve results should I keep? Each agent ships its own panel, its own task list, its own session history, so the review step is scattered across tabs and lost between restarts.
This is the right moment because the pain is fresh and the audience is identifiable. The people commenting on that thread are the people who already feel it, and they describe their own workflow in public. A local-first inbox needs no backend, no account and no agent SDK — it can be built and shipped in an afternoon, which means the idea can be tested against real traffic today instead of next quarter.
How It Works
Paste your agent output
Copy the raw log or JSON lines from Claude Code, Codex or Cursor and drop them into the inbox. A small parser reads the agent name, the task description, the status and the timestamp, so you never have to change an agent config or install an SDK.
Review one card at a time
Every background task becomes a single card with three buttons: accept, ignore, or defer. Counters at the top track how many items are still pending, so you always know how much unreviewed agent work is sitting in the queue.
Export what you accepted
When the queue is clear, export the accepted items as a Markdown checklist. Paste it straight into a pull request description or a commit message, and the decisions your agents made become part of the project history instead of disappearing.
What You Get
Zero integration
No SDK, no daemon, no config file to edit. You paste text and cards appear. That means you can try it against a real backlog of agent output in under a minute, without touching any of the agents you already have running.
Local-first by default
Every card, every decision and every export lives in your browser's local storage. Nothing is uploaded, no account is created, and no source code leaves your machine, which matters when the agent output contains private repository content.
Cross-agent by design
The parser is not tied to one vendor. Logs from Claude Code, Codex and Cursor all land in the same queue, so switching tools or running several at once does not fragment