agentic AI · developer tools · observability · tool · 2026
unwind
See your Claude Code and Codex sessions in one place: what each one said and what it cost. For Claude Code, agent calls appear as a live tree, so you can see who called whom and whether each one is still working.
Why I built it
I built unwind for two reasons. The first was a practical need: I couldn't see what my other project, callstack, was doing. The second was curiosity about cost.
Reason 1
I needed to see callstack's call trees
I was building callstack, which lets Claude Code call another full Claude Code session the way a program calls a function. That session does a big piece of work and returns its result. It can also call sessions of its own, so a single task can grow into a deep tree of calls. Claude Code showed all of these as one flat list of sessions. I wanted to see the tree.
Reason 2
How much am I really using?
I was paying $200 a month for my Claude subscription. I'd heard that subscriptions give you far more in tokens than you pay for, and I wanted to know if that was true for me. It was. Every month I was using billions of tokens, worth thousands of dollars at API prices.
The idea behind callstack
Subagents usually work in parallel and send results back loosely, so the agent that started them can miss something. callstack brings the discipline of function calls to agents instead. Each call runs to completion, its result comes back in a predictable way, and the caller collects every result before it moves on. Calls can nest as deep as the work needs, so full Claude Code sessions can take on large, complex tasks together. At the time, Claude Code had nothing like this.
unwind made callstack visible. I could watch calls appear as they were made and check that they behaved correctly: which calls were running, which had started calls of their own, which were waiting for their children, and which had finished.
Later: Codex too
I also started using OpenAI's Codex heavily, often on the same projects as Claude Code. I didn't want to check two places, so I added Codex support. Since version 0.5.0, unwind shows Claude Code and Codex sessions for a project together.
What it does
Run unwind serve in any project folder, and unwind opens a browser view of every Claude Code and Codex session that ran there, including months of history. Sessions started by callstack calls appear nested under the session that called them, not as a flat list. You don't install hooks or change how you use either tool.
For Codex sessions, unwind shows messages, tool calls, reasoning summaries, and token cost. Call trees and live status are available for Claude Code sessions.
Everything stays on your computer. unwind sends no data anywhere.
The whole multi-agent run as one tree
Subagents and callstack calls appear as connected cards, nested to any depth. Select any card to read that agent's conversation.
Know at a glance if it's busy, stuck, or done
New calls appear as they're made. Each card shows whether the call is running, has started calls of its own, is waiting for its children, or has finished. If an agent is waiting on you, that shows up all the way back to the root. Every card also shows what that session spent and what its whole branch spent.
Read any agent's conversation
Open a session to see each message, tool call, and reasoning step in order, with helpers nested where they started. It's the fastest way to find out why an agent stalled or failed.
Monthly spend across every project
This is the question that started unwind. The Reports view, or unwind usage report, shows a month's total tokens and their cost, then breaks both down by project. You can see Claude Code and Codex together or one at a time. Costs are estimates at each provider's API prices, which makes it easy to see what your subscription is really worth.
Status
Where it stands
unwind is stable and free to use. I built the first version in about six weeks in spring 2026 and shipped 18 releases to PyPI along the way. Version 0.5.0, released in October 2026, added Codex support.
It's now in maintenance. It does what I built it to do, so I'm not planning major new features. I'll keep it working as Claude Code and Codex change.
Timeline
- Apr 2026First public release, 0.1.0
- May 20260.2 to 0.4: twelve releases in four weeks
- Jun 20260.4.5, the last release before Codex support
- Oct 20260.5.0 adds Codex sessions
Open source and how to try it
unwind is open source under the MIT license. The code is on GitHub, and the package is on PyPI. You need Python 3.11 or later and a project you've used with Claude Code or Codex.
$ pip install unwind-labs $ cd path/to/your/project $ unwind serve # opens the view in your browser
unwind project | List and find the projects unwind knows about. |
unwind session | List sessions in a project, inspect one, or print its tree. |
unwind messages | Print or search a session's messages. |
unwind task | Explore the combined tree of calls and subagents. |
Add --harness claude or --harness codex to list sessions from one tool.
What you can use it for
Find the branch that blew the budget
When a run costs more than expected, follow the costs down the tree to the agent that spent it.
Check on long runs
See whether a run that's taking a while is busy, stuck, or waiting on you, without scrolling through a transcript.
See what your subscription is worth
Find out how many tokens you use each month, what they would cost at API prices, and which projects use the most.
Let Claude Code search its own history
Every command can return JSON, so Claude Code can use unwind as a tool. Ask it to summarize what the last subagent did, or to find the earlier session where you solved the same bug.
Debug multi-agent workflows
If you build with subagents or callstack, see how a run actually branched and read exactly what each agent was told.
Use Claude Code and Codex side by side
If you switch between both on the same project, see all their sessions in one list.
Explain agent runs to others
Export the tree as an image to show a team how a multi-agent run unfolded.
Build your own deep call stacks
Pair unwind with callstack to try nested agent calls and watch every level run.