Claudefolio
A portfolio research platform any AI agent can hook into. Data checks, agent reviews and a checked order sheet. A human approves every trade.
What we've been building
A friend asked what I've been up to. This is the answer, with pictures.
- 7,648 commits across seven projects, most of them since June 2026.
- Claudefolio, a portfolio research platform for AI agents, is the biggest: 5,854 commits.
- Most AI tokens are re-reads of saved context, not new writing.
Commits per month, seven projects
Nov–MayJune–Oct
Chart data
| Month | Commits |
|---|---|
| 2025-11 | 5 |
| 2025-12 | 46 |
| 2026-01 | 26 |
| 2026-02 | 11 |
| 2026-03 | 6 |
| 2026-04 | 52 |
| 2026-05 | 12 |
| 2026-06 | 653 |
| 2026-07 | 890 |
| 2026-08 | 1373 |
| 2026-09 | 2348 |
| 2026-10 | 2226 |
Projects
Agent Host
The server that lets ChatGPT and Claude run tools and long jobs on my machine.
crss
Keeps Claude Code sessions alive on a server, reachable from a phone.
Wallet Radar
Compares 161 crypto wallets, cards and ramps for developers, with scores you can rerun.
DispatchBench
Tests which AI model should do each coding job, and at what cost.
Questrade UI adapter
Lets AI agents trade on Questrade through a browser, behind the human approval gate. Aims to match our IBKR MCP. Work in progress.
This site
Where the lab shows its work.
Timeline
— Wallet Radar's first commit.
— This site goes up.
— crss: Claude Code sessions that survive on a server.
— Claudefolio starts.
— Agent Host: ChatGPT gets tools on the server.
— Agent Host simplified to long-running tasks.
— DispatchBench starts. Its results set who builds what.
— Claudefolio review workflow lands.
— Agent Host learns resumable file transfers.
— Order sheets show side-by-side decisions for the human.
— Token-use audit across local session logs.
The AI tokens behind it
I count tokens, the units AI models read and write. Most are cheap re-reads.
Re-readsCache writesFresh inputOutput
Chart data
| Tool | Category | Tokens | Share |
|---|---|---|---|
| Claude | Cache reads | 19.38 B | 96.6% |
| Claude | Cache writes | 0.57 B | 2.8% |
| Claude | Output | 0.12 B | 0.6% |
| Claude | Fresh input | 0.002 B | under 0.1% |
| Claude | Total | 20.08 billion tokens | |
| Codex | Cached input | 257.37 B | 97.6% |
| Codex | Fresh input | 5.58 B | 2.1% |
| Codex | Output | 0.64 B | 0.2% |
| Codex | Total | 263.59 billion tokens |
- A token is about three-quarters of a word.
- Re-reads cost about a twentieth of fresh input at list prices.
- I pay flat monthly plans, so this is usage, not a bill.
How these numbers were made
- Commits:
git logon each project's main checkout, counted by author month, duplicates removed. Agent Host uses its newest local branch. - Claude tokens: 3,473 local session logs, deduplicated per message, audited 2026-10-29.
- Codex tokens: 3,818 session logs with final counters, audited 2026-10-29. A later check flags these totals as disputed, so treat them as rough.
- Not counted: ChatGPT chat usage, which has no token meter we can read, and small Grok, Devin and Kimi runs.
- No invoice-backed dollar total exists yet. That is why the chart shows tokens.
Limits
- Commit counts include merge commits and agent commits. They measure activity, not value.
- Early months are partial: some old session logs are gone.
- One snapshot in time: 2026-10-31.