Show HN: Hiver – Chrome DevTools for AI Agents (hiver.sh)

Hey HN. I’m Emmanuel, and I built Hiver.

I kept running into the same problem while building agents: when something failed, the logs showed isolated commands, but not the full sequence of what the agent saw and changed. The browser was in one place, terminal output somewhere else, file changes in another system, and network-policy decisions buried in infrastructure logs.

Hiver puts all of that on one synchronized timeline: model activity, tool calls, terminal output, browser state, filesystem operations, network egress, resource usage, and human approvals.

The demo on the homepage is a real interactive replay, not a video. It shows an agent triaging email, checking a calendar in a separate browser sandbox, scheduling a call, generating a report, and sending it. You can pause, scrub, change playback speed, filter events, and inspect the state of the terminal, browser, files, and network at any point.

Hiver also provides the runtime that produces the trace. Agents run in isolated sandboxes where file access and network egress can be governed and audited. A workflow can span multiple sandboxes—for example, one running Claude Code and another running Chrome—while still appearing as a single end-to-end execution.

There are already many products that give agents a sandbox. Hiver’s bet is that execution, policy enforcement, and debugging should be one system. The same runtime that allows or denies an action emits the event you later inspect, so a failed request appears in context alongside the agent’s reasoning and the surrounding browser and file activity.

It is agent-agnostic. You can bring Claude Code, Codex, Copilot, an agent SDK, or your own loop. The project is open source and can run locally or in your own cloud.

I’d especially value feedback on two things:

Does the replay make it materially easier to understand what the agent did?

What would you need before trusting this for production agent workflows?

I’ll be around to answer technical questions and would be happy to discuss the runtime, isolation model, browser integration, or how the replay is recorded.