Doug Stewart

[Doug Stewart @dstewartzz]

Does anybody know a way to track token usage in Claude Co-Work on a per-project basis so those costs can be billed back to a client? I don't think the analytics API is going to make that possible.

[Doug Stewart @dstewartzz]

@businessbarista Ai native

[Doug Stewart @dstewartzz]

@mzahidtech @cursor_ai Could this work for knowledge work tasks too? Sure would be cool if it could.

[Doug Stewart @dstewartzz]

@seandsweeney Sean, not sure this is your problem, but when I first signed in this afternoon, I noticed that it had rearranged all of my chats by project rather than date. Any chance you just have to filter them by date and they’ll come back?

[Doug Stewart @dstewartzz]

@ansubkhan @openclaw @karpathy Thanks you for building this. Realized this weekend that my Wiki needed a way for our attorneys to visualize it. Very helpful.

[Doug Stewart @dstewartzz]

@trq212 Not sure I’m a mutual but definitely interested.

[Doug Stewart @dstewartzz]

@jerryjliu0 This is fantastic for attorneys, thanks!

[Doug Stewart @dstewartzz]

@maxhodak_ Just dive into Code. I’m a non-tech person, took one beginning CC class with @every, and have never gone back to Chat or Co -Work. It does all of the coding.

[Doug Stewart @dstewartzz]

@dansemperepico Even worse, things I’m setting up in Code for my firm do not seem to work in Cowork as I assumed they would.

[Doug Stewart @dstewartzz]

@BitGrateful I’m just using the expanded summary from granola through the MCP into case notes, as opposed to the entire transcript.

[Doug Stewart @dstewartzz]

@stockthoughts81 I’ve tried to work with OneNote through the MCP but Claude struggles with it mightily. Claude suggested MD files instead. I read mine with Obsidian. I think a MD structure, Obsidian, or Notion might work for you instead.

[Doug Stewart @dstewartzz]

@strikerglows @amorriscode I was wondering the same thing and noticed in another thread that the response was that they were pretty much the same thing.

[Doug Stewart @dstewartzz]

Thanks Chris, would be nice, but not a deal-killer, to provide access to shared team mailbox so I can get firm-wide visibility through Claude without going through each individual user. I appreciate your responsiveness and have been very pleased with the MCP access. Using it and O365 for my entire firm knowledge base management.

[Doug Stewart @dstewartzz]

The CSV output is designed for accounting import:
date, case_number, description, session_id, duration, input_tokens, output_tokens, cache_write, cache_read, cost_usd
/bill report 2026-03 gives you a grouped monthly summary by case in a csv file that you can drop into your accounting program.

I welcome suggestions or constructive criticism, but bear in mind that I am an attorney, not a computer programmer. Claude Code basically took my suggestion, created the skill, pushed it to GitHub, and helped me write a thread that would clearly explain what it created for me. I only wish it worked for Cowork or Chat, but it does not and Claude reports that there's no mechanism available for that — Anthropic would need to build it into the platform.

[Doug Stewart @dstewartzz]

After watching and reading on X for months I finally took a class on Claude Code from @every and have been having a blast "coding" (well, repeatedly pushing the accept button) ever since. It didn't take me long to figure out that my new hobby was taking a lot of tokens so I had Claude build an open-source billing tracker for Claude Code (not Chat or Cowork) that lets professionals recover appropriate AI costs as client expenses. https://t.co/xfWnp0FXk6

/bill, client code, and description at session start. When the session ends, a SessionEnd hook automatically parses the JSONL transcript, extracts per-message token usage (input, output, cache read/write), applies Anthropic's published per-model rates, and appends a row to a CSV.

Three files, zero npm dependencies, pure Node.js built-ins. Works with mixed-model sessions (Opus/Sonnet/Haiku) and captures subagent costs. Monthly report generation built in.

[Doug Stewart @dstewartzz]

It handles edge cases: mixed-model sessions (Claude spawns Haiku subagents for searches — those get different rates), cache write 5-min vs 1-hour tiers, subagent transcript files in nested directories. All from parsing what Claude Code already logs — no API proxy needed.

[Doug Stewart @dstewartzz]

@DivyanshT91162 Thank you for this post. Question from a non-technical knowledge worker who makes Claude read a lot of context, why is it more efficient to pass work to sub-agents? Intuitively it seems that spreading out the work would take a similar amount of tokens? Thanks again.

[Doug Stewart @dstewartzz]

@osmarks1 @Dorialexander You missed a very significant part of the article. What experienced lawyers add is judgment. I spent the better part of my day yesterday working through issues with Claude that required the judgment I have earned in 30 years of experience. That’s what the human lawyer adds.

[Doug Stewart @dstewartzz]

@StephanieKelton My wife says blue sky