Deterministic browser recordings an agent can debug at any execution point.
replay.io ↗repo ↗docs ↗Replay MCP tool reference ↗How does time travel work ↗Pricing ↗replay-cli (npm replayio) ↗Replay Chromium fork ↗
Different domain (debugging arbitrary JS runs in the cloud) but the best existing model of what agents want to ask a history.
Borrow from it: take a specific idea, API or format.
Replay is the strongest proof that agents debug better with time travel than with code reading: it records a whole browser's runtime inputs so the run is exactly reproducible, then hands the agent ~30 MCP tools addressed by execution point — evaluate any expression at any moment, read the stack, see which lines ran, ask why a React fiber re-rendered, read Redux state at a dispatch by path, screenshot the page at a timestamp, and send the human a link to that exact moment. It solved the hard problem (determinism for arbitrary, effectful JavaScript, not just pure Elm-style apps) and the agent-ergonomics problem (summary-first modes, every result carries points the next tool accepts). A fan would say: this is what 'agent-addressable time travel' means in practice, it already works with Claude Code today, and anything scene builds should match its readback depth.
| UI as data | 0 scene 5 | Records arbitrary web apps; the UI stays code. |
|---|---|---|
| Agent can drive it | 3 scene 5 | Rich MCP, but over recordings, not a live instance. |
| Agent can see it | 5 scene 5 | Structured state at any point plus screenshots at any timestamp. |
| Small, targeted edits | 1 scene 5 | Only retroactive logpoints/evaluation. |
| Time travel | 4 scene 0 | Agent can read history and seek to any execution point; no fork or modified replay. |
| Live data rate | 0 scene 5 | Post hoc only. |
| Terminal native | 0 scene 5 | |
| Pixel graphics | 0 scene 4 | |
| Teaching | 2 scene 3 | GetPointLink and shareable recordings help explain bugs; no tour features. |
| Maturity | 3 scene 1 | |
| Openness | 1 scene 4 | Open clients (MPL/BSD), proprietary hosted core and MCP. |
Replay.io scene now scene planned
Hosted MCP server (dispatch.replay.io/nut/mcp) with ~30 tools, plus agent skills (replay-mcp, replay-cli, replay-playwright). The agent records via `npx replayio record <url>` (a human reproduces, then closes the browser) or a Playwright run, then investigates by recording ID.
Low: add the MCP server and skills to Claude Code/Cursor/Codex; requires a Replay account (free tier 25 credits/month) and uploading recordings to Replay's cloud.
Very strong, structured: RecordingOverview, ConsoleMessages, UserInteractions, NetworkRequest, LocalStorage, DescribePoint (variables, executed lines, dependency chain), Evaluate, GetStack, ReactComponentTree at a time, ReduxActions action-state by path, Screenshot at a timestamp, InspectElement.
No edits to the recorded program; the only 'writes' are retroactive logpoints and expression evaluation at a point.
Seek and read: any execution point or timestamp, for state, stack, DOM, screenshot, React tree, store state. No fork, no replay with modified inputs, no live control.
The most capable agent-facing read-only time travel found: an agent can interrogate any moment of a past run, but cannot change or branch it.
Runtime replay, not session replay: a modified browser (Chromium fork, also a Node runtime) records the nondeterministic inputs to the JS engine (network, timers, events), so the recording can be re-executed exactly in the cloud. Any 'execution point' can then be reconstructed on demand: evaluate expressions, read stacks and variables, add logpoints retroactively, take screenshots, walk the React fiber tree. Recordings upload to Replay's backend; humans use Replay DevTools in the browser, agents use the hosted Replay MCP server, both against the same recording ID.
No published recording-overhead or replay-throughput numbers found. Analysis is minutes-scale on first open; inherently post hoc.
Respected among frontend framework authors; modest CLI usage. Stars are for replayio/devtools. first_release is the npm replayio package creation date, used as a proxy.
Used by: Vercel / Next.js team (testimonial: 'Next.js 13.4 wouldn't have been possible without Replay', Tim Neutkens), Dan Abramov (testimonial)
Actively developed as a company product, but the open repos are peripheral and the company's focus has shifted from devtools to autonomous QA.
Releases: Continuous SaaS deploys; CLI 1.9.1 on 2026-09-01. Product Hunt launches 2026-07-20, 2026-08-17, 2026-09-08 (Replay QA). · Contributors: replayio/devtools 65 (GitHub contributors API) · Recent: replayio/devtools: ~6 commits in last 90 days (open-source frontend mostly quiet); Chromium fork pushed 2026-10-08; MCP docs current. · Bus factor: medium: company-run, closed core; viability tied to one startup's pivot to QA
Agent-addressable history: read state at a past point, screenshot at a time, shareable link to a moment. scene's planned state_at/screen_at/diff are Replay's DescribePoint/Screenshot for a terminal UI.
The agent-facing tool shape for history inspection (summary-first, point-addressed, drill-down modes). Replay has already iterated on what an agent needs to read from a past run; scene should copy the ergonomics, not invent them.
Replay cannot fork, replay with changes, or drive a live UI; it is web-only, cloud-hosted and closed. scene gives an agent live control plus seek/fork on a local, open, terminal UI whose determinism comes from architecture rather than runtime recording.
| 1–7 | home · features · use cases · docs · changelog · under the hood · blog |
| t | light or dark |
| j k | scroll |
| g G | top · bottom |
| ? Esc | this · close |