Boston AI Week · 22 October 2026
Tokenome, end to end
John O'Neil and Sid Probstein gave a training talk on Tokenome: why it exists, installing it from scratch, pointing at a line of code and reading the conversation that produced it, search, the decision journal, agents querying their own history over MCP, what the evaluation can and cannot claim, and the self-hosted team server. The eight sections below are that talk, in the order it was given, twenty one minutes in all.
Sound off, or short of time? Each section has the whole answer in
writing above its player, so you can read it instead. The project shown,
acme-payments, and all of its conversations are sample data; the software
is not.
Prefer it unsplit? The whole talk runs 37 minutes with chapters, on YouTube. The eight sections below are the same recording, so you need one or the other, not both.
The problem
Why we built Tokenome
The reasoning behind your code is trapped in chat history. Git tells you what changed, when, and who changed it. It cannot tell you why, because the reason was in the conversation that produced the change, and that conversation is in a different tool, unlinked and invisible. When the context window goes, so does the why.
Install
Installing it, live
One command, and three things on your PATH. There is nothing to configure on first run: it finds the tools you already use, indexes what they have written, and starts answering. Nothing leaves your machine. No account. No cloud embeddings.
Code provenance
Point at a line. Get the conversation.
This is the product in two minutes. Point at line 43 and Tokenome brings back the conversation that wrote it. Point at line 31 and you get the change that was made and then undone, with the reason it was reversed, which is the part no commit records. The honest limit is stated in the section itself: if nobody ever discussed the line, you get nothing back.
Search
Search in your own words
Keyword and meaning in the same query, across every tool you use: Claude Code, Claude Desktop, ChatGPT and Gemini. The index is on your machine and the embeddings run on your own device, so searching your history does not send it anywhere.
The journal
Decisions, with the receipts
Once a day, in plain markdown, Tokenome writes down what was decided and quotes the line it came from. If a claim cannot be quoted, it is not written down. A decision that was later reversed reads as a reversal rather than quietly disappearing, so the record matches what actually happened.
Agents and MCP
Your agent asks its own history
Tokenome ships an MCP server, so the agent queries the history itself instead of waiting for you to paste something in. In forty controlled sessions, routing agent context this way used 38% fewer context tokens with no measurable change in answer quality. That is context, not money: dollar cost fell 22% in the same arm.
The evaluation
What we can claim, and what we cannot
Forty controlled sessions, ten questions asked four ways, and the honest reading of the result. The baseline arm was repeated three times with nothing changed, so the noise floor is known to be 5.4% and the reductions sit well outside it. The section also covers what the evaluation does not support, which is the part that usually goes missing.
Tokenome Team · in beta
One search across everyone
One query covers your own history and everyone's opted-in projects, and every team hit carries the name of the person it came from. You pick which projects to share, and secrets are redacted on the laptop before anything leaves it. The server is yours: one container image, on your own infrastructure. We do not run one for you.
Team is in beta, and free while the beta runs. No end date is set.