- SPECIMEN
AgentPlay
- FILED AS
Open-source, local replay viewer that maps coding-agent sessions step by step
- INTAKE DATE
- 2026-10-20
- CLASS
- Coding
- METHOD
- 4 pages scraped, 3 search passes, 6 community sources
It covers file touches, shell commands, MCP tools, subagents and error classes on one step timeline across several agents, though very large sessions hit hard size caps. THE RUNDOWN
A single uvx or pipx command builds the maps and opens a local viewer, with no account or API key to set up first. ZERO TO RUNNING
The full tool is free and open source, so the output you get per dollar has no ceiling to argue about. THE DAMAGE
Read on 2026-10-20, the GitHub repo shows 261 stars; 0 releases in the last 90 days, a small start for a launch Entelligence announced on Reddit only days earlier. github.com
Read by Koda: Entelligence-AI/agentplay on GitHub, read 2026-10-20: 261 stars; 0 releases in the last 90 days. github.com
Koda Score = weighted blend: capability 35, ease 20, value 25, momentum 20.
AgentPlay reads transcripts your agents have already written, runs entirely on your own machine under Apache-2.0, and pins the moments worth inspecting to a timeline: retries, context compactions, model switches and opened PRs. Trying it costs one command and no account, so the only real question is whether you review agent runs often enough to keep opening it. Developers who regularly untangle long agent sessions should install it now; teams hoping for shared review workflows will find those parked in a hosted edition.
AgentPlay turns the coding-agent sessions sitting on your laptop into a replayable map, so you can scrub to the exact step where Claude Code or Codex took a wrong turn.
Captured: entelligence.ai/agentplay on 2026-10-20.
Read this first.
3 LINESTry it with uvx agentplay; Entelligence says there is no account or API key involved.
Jump to the Retry and Error markers first, since that is where a derailed run usually shows its hand.
Export a known-good session to hand a teammate a concrete example of agent-assisted work done well.
What it can do.
7 CAPABILITIES- Interactive session map
- Lays out the repositories, files, shell commands, MCP tools, models and subagents an agent touched as one clickable graph.
- Signal markers
- Flags repeated failed commands, slow tool calls, compactions, model switches and opened PRs directly on the replay timeline.
- Compaction view
- Shows when context was compressed with token counts before and after; the homepage demo drops from 168.4K to 23.1K tokens.
- Error classes
- Groups failures by type, from a shell non-zero exit to an overloaded model API, and notes when a call was retried.
- Export and import
- Writes a session map to .json or .json.gz so someone on another machine can add it with agentplay import.
- Incremental rebuilds
- Later runs rebuild only the sessions whose transcript changed.
- Agent detection
- The agentplay agents command reports which supported coding agents it found on the machine.
Ways to run it.
3 PLAYS- Post-mortem a derailed run
- Open the failed session, scrub to the first Error or Retry marker, and read the step just before it. Add one line to your instructions file that rules out that move.WHODeveloper using Claude Code or CodexPAYOFFA prompt fix aimed at the actual wrong turn, written from evidence.
- Compaction check
- Whenever output quality sags late in a session, look for a compaction or model switch marker sitting right before the drop; if one is there, split that kind of task into shorter sessions from now on.WHODeveloper running long agent sessionsPAYOFFA simple rule for when to start a fresh session.
- Onboarding reel
- Pick two clean sessions, export them, and have new hires import and replay them during their first week.WHOTeam leadPAYOFFNew developers see what good agent-assisted work looks like, step by step.
What it costs.
1 TIER- Verdict on the price
- Free and local is hard to beat; the open question is the hosted team edition, whose price did not appear on the pages read.
The street's view.
1 QUOTEDToo new for a real street reaction: past Entelligence's own Reddit launch posts, a YouTube short and a YouTube explainer are the only independent takes so far.
A YouTube tester tried the Claude Code session replay on their own history and pointed out that the Review tab shows sample data.
MixedWhere it breaks.
4 LIMITATIONSA single map stops at 20,000 events, so marathon sessions may not render in full.
Transcripts larger than 512 MB are skipped outright.
Entelligence labels Windows support as less tested.
It shows you where things went sideways; turning that into a better prompt is still your job.
What else to weigh.
3 ALTERNATIVES- claude-replay
- Converts Claude Code, Cursor, Codex, Gemini and OpenCode sessions into self-contained, embeddable HTML replays.PICK IT WHENYou want to post a replay on a page or in docs as a single HTML file.
- Coding Agent Session Search (CASS)
- Searches coding-agent histories across tools from a terminal UI.PICK IT WHENSearching across many past sessions matters more to you than replaying one, and a TUI that one comparison calls clunky does not bother you.
- Spotlight by Backplanes
- Session reports for Claude Code and Codex focused on improving future sessions.PICK IT WHENYou want recurring written reports on agent behavior over an interactive map.
First hour.
4 STEPSConfirm Python 3.11 or newer is installed.
Run uvx agentplay (or pipx run agentplay, or pip install agentplay) in a terminal.
Let it build maps for the sessions it finds; it starts a local server and opens the replay in your browser.
Run agentplay agents if a session is missing, to see which supported agents were detected.
The questions people ask.
4 ANSWERS- Does my session data leave my machine?
- Entelligence says agentplay runs locally with no telemetry or remote data usage. Its suggested setup prompt also tells your agent to avoid changing agent settings and to upload nothing.
- Which coding agents can it read?
- Entelligence's Reddit launch post names Codex, Claude Code, Cursor and Copilot sessions. The homepage says it supports multiple agents generally.
- How do I share a session with a teammate?
- Use agentplay export to write the map file, send it over, and the teammate adds it with agentplay import. Nothing needs to be hosted for this.
- Is there a paid version?
- The GitHub README says Entelligence offers a hosted team edition that adds review comments traced back to the session edit that caused them. The pages read showed no price for it.
The call.
SPECIMEN 0256 CLOSEDKeep it installed for the next time an agent loops on a failing test. Ten minutes in the replay will usually show which line of your instructions file needs rewriting.
Field research: 4 pages scraped · 3 search passes · 6 community sources. Reviewed by the Koda desk on 2026-10-20.