agentplay: replay a coding agent session as a map and timeline

Your coding agent’s transcripts already record everything: which files it read before editing, which commands it retried, which MCP servers failed, where the context window filled up, which subagents came back empty. That data sits in ~/.claude, ~/.codex, ~/.cursor and similar folders, and almost nobody looks at it. agentplay, an Apache-2.0 Python tool from Entelligence, turns each session into a radial map plus a playable timeline so you can see what the agent actually did. The repository was created on October 14, 2026, has 316 stars, and was still receiving pushes today.

Core features

• A radial session map: repos, directories and files (split into seen, read and edited), shell verbs, MCP servers and tools, skills, subagents, web hosts, error classes, models, and the outputs it produced — commits, PRs and tickets — laid out as one tree.
• A timeline you can play: 1x, 4x or 16x playback, switching between a step axis and wall-clock time. Lanes cover prompts, errors, delays, outputs, edits, exploration, MCP, skills, tools, model, cost, tokens, subagents and context size, with compactions marked.
• Tasks and an agent lens: a session is split into tasks, usually one per prompt; pick a task, a model or a single subagent and the whole explorer re-scopes to it.
• Error ranking: the Errors tab groups every error by class and ranks them by the longest run of the same failing command; click a row to seek the replay to that moment.
• Nine agents out of the box: Claude Code, Codex (CLI and IDE), Cursor, GitHub Copilot CLI, Copilot in VS Code, OpenCode, Kiro, Gemini CLI and Antigravity CLI. Claude Code has the highest fidelity — read straight from JSONL including hooks, compactions, stop reasons and latency.
• No dependencies, no uploads: nothing beyond the Python standard library at runtime, and uvx agentplay@latest runs it in one command. Session data never leaves your machine, and a map contains no prompt text, tool inputs or outputs, file contents or command arguments — a shell command is recorded as its verb only (git log, never git log --author=…).

Typical use cases

• Debugging a session that went wrong: when the agent edits the wrong file, the map’s read list tells you what it had actually looked at, far faster than scrolling thousands of lines of JSONL.
• Chasing tool and MCP failures: which MCP server kept failing, which command was retried a dozen times — the Errors panel ranks them for you.
• Seeing where the money went: the token, cost and context-size lanes plus compaction markers show exactly which step blew up the context window.
• Sharing a session: agentplay export writes a single map file, and rebuilding with --path-mode dirs_only keeps only directory-level paths so it is safe to hand to a colleague.
• Fitting existing workflows: for reviewing the code the agent produced, pair it with PR-Agent; for a ready-made Claude Code engineering setup, see ECC.

Quick start

1. You need Python 3.11 or newer, then run uvx agentplay@latest (or pipx run agentplay, or pip install agentplay).
2. It finds the agent sessions already on your machine, builds redacted maps incrementally, and opens a local viewer at 127.0.0.1:7317. Later runs only rebuild sessions whose transcript changed.
3. Useful subcommands: agentplay build builds maps only, list shows built sessions, export <session> writes one map, import adds a teammate’s maps, agents reports which agents were found.
4. To keep it fully offline, set AGENTPLAY_NO_NETWORK=1 before launching; that also hides the viewer’s optional email signup.

Summary

agentplay’s value is turning “what did the agent do” from a log into something readable: the map answers what it touched, the timeline answers how it got there, and the error ranking answers where it got stuck. Apache-2.0, no runtime dependencies, and data that stays local make it easier to try than most observability products. The limits are stated plainly too: outside Claude Code, subagents do not get their own lanes; Cursor records no tokens; and an exported map needs its path mode checked before it is shared publicly.