# agentplay

> Open-source local tool that replays coding agent sessions as an interactive map and timeline.

agentplay is an open-source session replay tool from Entelligence that reads the coding-agent transcripts already stored on a developer's machine and turns them into interactive maps and timelines. It supports Claude Code, Codex, Cursor, GitHub Copilot, OpenCode and Kiro, and runs locally with no account or API key. It is licensed under Apache 2.0.

## What It Is

agentplay is a local session explorer for coding agents. Running a single command finds agent sessions on disk, builds a redacted replay of each one, and opens a viewer served on 127.0.0.1. Each session becomes a radial map of the repos, files, shell commands, MCP servers and tools, subagents, models, errors and outputs (commits, PRs, tickets) it touched, plus a timeline that can be played back at 1x, 4x or 16x.

## What You Can Inspect

The replay is meant to answer what the agent read before editing, which commands failed or were retried, where the context window filled up and was compacted, when models switched, and which PRs or commits the session produced. The viewer supports keyboard navigation to jump between edits, errors, delays, outputs and retries, and scoping by task, model or a single subagent. An Errors tab groups failures by class.

## Privacy and Sharing

The project states that session data never leaves the machine and that maps exclude prompt text, assistant text, tool inputs and outputs, file contents and command arguments. Path modes (relative, dirs_only, hashed) control how file paths are stored. A single session can be exported to a .json or .json.gz map file and imported by a teammate. The only network call is an optional email signup, which can be disabled with an environment variable.

## Setup Path

It requires Python 3.11 or newer and has no runtime dependencies. It can be run with uvx, pipx, or installed with pip, and the project provides an install.md written for coding agents. Costs shown are estimates from public list prices, and Cursor sessions carry no token or cost figures. Maps are capped at 20,000 events, and Windows is less tested than macOS and Linux.

## Features
- Interactive radial map of files, tools, MCP servers, subagents, models and outputs
- Playback timeline at 1x, 4x and 16x with step or wall-clock axis
- Markers for retries, delays, errors, context compactions and model switches
- Detection of commits, PRs and tickets produced by a session
- Scoping by task, model or subagent
- Redacted maps with configurable path modes
- Export and import of session maps as .json or .json.gz
- Incremental builds that only rebuild changed sessions
- Open agentplay/v1 session map format with JSON Schema
- Runs locally with no account or API key

## Integrations
Claude Code, Codex, Cursor, GitHub Copilot CLI, GitHub Copilot in VS Code, OpenCode, Kiro

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, VSC_EXTENSION, CLI

## Pricing
Open Source

## Links
- Website: https://entelligence.ai/agentplay
- Documentation: https://github.com/Entelligence-AI/agentplay#readme
- Repository: https://github.com/Entelligence-AI/agentplay
- EveryDev.ai: https://www.everydev.ai/tools/agentplay