What are thinking blocks, and why would I read them?

Modern models reason before they answer, and that reasoning lands in the session transcript even though your terminal shows a spinner. Reading it back is the closest you can get to asking the model what it was doing. When a session makes a strange decision, the thinking around that turn usually names the assumption that caused it.

Where do my transcripts actually go?

The watcher on your machine tails the transcript files your agents already write and streams them to a server instance dedicated to your organization, run by Kurrent. Sessions land in an append only event store, and the dashboard and search read from that. The client is source available, so what leaves your machine is inspectable. Read their security page before pointing it at a codebase you are not allowed to share.

Does it slow the agent down or change how it behaves?

No. The agent writes its transcript the way it always has, and Capacitor reads that file as it grows. Your prompts, your tools, and your workflow stay exactly as they are.

Which agents does it cover?

Setup detects what is installed on the machine and wires up what it finds. Our own workspace records Claude Code and Codex sessions side by side, and both are visible in the video. The vendor describes support for coding agents generally, so check their documentation for anything beyond those two.

The free plan carries five seats and every feature, with no credit card, and you bring your own AI keys. Paid plans add more analysis inside the app. There is no affiliate link anywhere on this page. We run it because it is useful, and that is the whole relationship.

What would I actually do with months of session history?

Ask it questions. Which sessions burned the most tokens for the least code. What the agent was told before it made the change nobody can explain. Where the long sessions go wrong, and whether it is the same place every time. The record turns those from arguments into lookups.

Video guide

See what your coding agent was actually thinking.

Claude Code shows you a spinner, then a diff. The reasoning in between scrolls past once and is gone when the terminal closes. This guide sets up Kurrent Capacitor, a session recorder that keeps every session your coding agents run, transcript and thinking blocks included, in a workspace you can search. The next time a session burns an hour going nowhere, you can open it back up and find the exact turn where it drifted.

Watch the two minute versionOn Instagram. The steps below stand on their own if you would rather read than watch.

Setup, once: 10 min
Steps to follow: 6
Five seats, no card: Free
Session recorded: Every

Where this leaves you

One install on the machine your agents run on. Every session after that is kept, searchable, and readable end to end.

A workspace at your own address, on the pattern you.kcap.ai, holding every session your agents run from now on
The full transcript of each session, with the thinking blocks the terminal never shows you
Token figures on every session, so the expensive ones stand out in the list
A way to answer, after the fact, where a long session went off the rails and what you would change next time
The same record for everyone on the team, five seats on the free plan

Do it once

Six steps. This is the flow we ran for our own workspace, the one in the video, on August 29, 2026.

A machine that already runs Claude Code or Codex, with Node and npm on it. If Claude Code arrived through npm, you already have both.

01
Install the client from npm. One global package.
npm install -g @kurrent/kcap
The client side is source available, so you can read what you are about to run before you run it.
02
Run setup in the same terminal. A browser tab opens for the account step. If you are new, pick the sign up option at the bottom of the form.
kcap setup
03
Back in the terminal, name your organization and choose your workspace address. It follows the pattern you.kcap.ai.
The workspace in the video is ours. Yours is whatever you pick here, and the name is the hardest part of the whole setup.
04
Give it a few minutes. It provisions the workspace, then finds the coding agents installed on the machine and wires up their integrations on its own.
Nothing about how you work changes. A background watcher tails the transcript files your agents already write.
05
Start a coding session the way you always do. Then open your workspace in the browser, click into the session, and open the Transcript tab. The thinking blocks are in there in full, alongside every prompt and tool call.
Sessions, then your session, then Transcript
Overview, Events, and Trace sit beside the transcript. Agents, Insights, Flows, and Work items fill in as more sessions land.
06
Use it for the retro. When a session ships the wrong thing, read the thinking blocks around the turn where it drifted. That is where you find out whether the prompt, the context, or the plan was the problem.
This is the habit that makes the recorder worth having. The transcript is a record of how you and the machine work together, and it reads differently a day later.

Capacitor has paid plans that do more analysis inside the app. Everything on this page sits on the free tier.

Questions people ask

What are thinking blocks, and why would I read them?
Modern models reason before they answer, and that reasoning lands in the session transcript even though your terminal shows a spinner. Reading it back is the closest you can get to asking the model what it was doing. When a session makes a strange decision, the thinking around that turn usually names the assumption that caused it.
Where do my transcripts actually go?
The watcher on your machine tails the transcript files your agents already write and streams them to a server instance dedicated to your organization, run by Kurrent. Sessions land in an append only event store, and the dashboard and search read from that. The client is source available, so what leaves your machine is inspectable. Read their security page before pointing it at a codebase you are not allowed to share.
Does it slow the agent down or change how it behaves?
No. The agent writes its transcript the way it always has, and Capacitor reads that file as it grows. Your prompts, your tools, and your workflow stay exactly as they are.
Which agents does it cover?
Setup detects what is installed on the machine and wires up what it finds. Our own workspace records Claude Code and Codex sessions side by side, and both are visible in the video. The vendor describes support for coding agents generally, so check their documentation for anything beyond those two.
What does it cost?
The free plan carries five seats and every feature, with no credit card, and you bring your own AI keys. Paid plans add more analysis inside the app. There is no affiliate link anywhere on this page. We run it because it is useful, and that is the whole relationship.
What would I actually do with months of session history?
Ask it questions. Which sessions burned the most tokens for the least code. What the agent was told before it made the change nobody can explain. Where the long sessions go wrong, and whether it is the same place every time. The record turns those from arguments into lookups.

Why a business is writing about session recorders.

GraniteAI builds and runs AI systems for small and midsized businesses, out of Manchester, New Hampshire. When work starts running through AI agents, the session record becomes the paper trail of how the work actually happened, and tools like this one are how we audit our own. If your business leans on AI and nobody can say what the AI did last Tuesday, that is a problem we know how to fix.

See what we build Browse the other guides