I had 801 agent transcripts and no way to read them

Every time you run Claude Code it writes a transcript. Every subagent it spawns writes its own. They pile up in ~/.claude/projects as JSONL, one JSON object per line, and nobody ever looks at them.

I checked mine: 801 agent transcripts across 26 projects. The Armies of Gielinor gauntlet run alone left 114 of them across five sessions, 295 MB. I had spent an entire night orchestrating those agents and I had no idea what most of them actually did.

So I built something to read them. It’s an MCP server now — you install it and ask Claude about your own usage in plain language, and it answers out of your own transcripts.

Install it

/plugin marketplace add aaronorelup/claude-playback-lens
/plugin install playback-lens@claude-playback-lens

Or skip the plugin and add the MCP server on its own:

claude mcp add --scope user lens -- npx -y claude-playback-lens-mcp

Node 20 or newer. Five tools — lens_status, lens_sessions, lens_usage, lens_search, lens_session — all read-only. The source is at github.com/aaronorelup/claude-playback-lens, MIT.

It speaks stdio, and that one is deliberate rather than a default. The payload is my entire private history — every prompt I’ve typed, every file I’ve had open, every screenshot an agent took of my screen. It has no business crossing a network, so there isn’t one to cross. No account, no upload, no telemetry.

What it found

I pointed it at session two of the gauntlet: 80 agents, 17 workflows, 8 hours 39 minutes.

An agent died four hours in and I never noticed. maps, on Opus, state error. It ran 38 minutes, burned 147,264 output tokens and 11.3 million cache reads, and produced no result — about $22, the fourth most expensive agent in the whole run. The journal recorded 80 agents started and 79 results returned. That one-line gap was sitting there the entire time and I had nothing that would surface it.

One agent wrote 90,189 characters of reasoning against 479 characters of prose. It thought enormously hard and then fell over.

The loop actually converged. Findings raised by the critics went 380 in round one, 126 in round two, 0 in round three. That is the thing I most wanted to know when I wrote the gauntlet post, and at the time I could only assert it.

The part that isn’t in the transcript

I looked before I built, because a transcript viewer is exactly the kind of thing six people have already done. Six people have, and the good ones are genuinely good. None of them worked for what I wanted, and the reason is specific enough to be worth writing down.

When you run a Workflow, the harness writes the orchestration to a file the conversation never touches:

<session>/workflows/wf_<id>.json

Inside it is a workflowProgress array, and that array is the only place on disk where an agent’s phase, state and attempt count survive. Meanwhile the agent’s own JSONL is the only place its tool sequence, screenshots and reasoning live.

Neither file is useful without the other. A viewer that opens only the session finds 80 anonymous sidechains and shows you 80 rows called agent-a771062a38caac9c6. Joining them is the entire trick, and it’s about forty lines of code — it just requires knowing the second file is there.

No inference, ever

Every number is a recorded fact from the transcript, or arithmetic over recorded facts. Nothing is clustered, scored, guessed at, or summarized by a model.

The part of that I’d defend hardest is the smallest one: an unknown renders —, with the reason, and a real zero renders 0. They are never the same glyph, and a missing rate is never quietly $0. It is very easy to write a cost dashboard that turns “I don’t know” into a number, and the moment it does, you have lost the ability to tell which of its figures you are allowed to trust.

Two things out of one repo

The MCP server is the product. The same repo also carries a viewer — git clone, then:

node lens.mjs --serve

The viewer is in dev mode, and that’s a real caveat. It is a six-level drill-down that exposes everything the index records, censuses and reconciliation pages included. That is the right shape for debugging the index and the wrong shape for a casual look at last week’s spend. I built it first and I still use it, and it is not what I hoped it would be as a finished thing. A friendlier one is separate work I haven’t started.

The decision I’d keep

The MCP doesn’t reimplement any of the analysis. It also doesn’t call the viewer’s HTTP API over the network, which was the obvious shortcut.

It builds the same router the server builds, registers the same handlers against the same context, and dispatches synthetic requests through it with a capturing fake response object — about fifty lines of it. So the MCP and the browser get byte-identical JSON out of the same handler. There is no second implementation that can drift.

That wasn’t for elegance. The payload builders are closures inside the API’s body, and they close over the memo objects that de-duplicate a single message id across forked and resumed sessions. You can’t export them without hoisting them out of that closure — and a fork is precisely where a second implementation would start quietly disagreeing with the first. The parity check asserts exact integer totals against a store that contains a forked session.

The part I keep coming back to

None of this is analysis, and I think that’s the piece worth taking away from it. Every number in it was already sitting on my disk, written down by the tool while it worked, in a folder I had never opened. I didn’t build something that figures anything out. I built something that reads.

801 isn’t an unusual number, either. If you have been running agents for a few months you have your own pile, and you haven’t read yours either.