So I’m currently helping out on this project, and bro. The "Context Tax" is real.
My lead dev was trying to debug an auth issue, but he spent like 20 minutes just digging through old Slack threads, random .env backups, and five different README.md files just to remind himself how the legacy middleware was structured. Watching him copy-paste blocks of code into Claude just to "catch the AI up" was painful. 🥀
I told him I could automate that. He gave me that "sure kid, whatever" look. Fair.
So I set up Trayce. It’s this local-first desktop app that indexes your entire project directory and local assets into a knowledge graph. What it does—it uses MCP (Model Context Protocol) to basically give the AI a direct pipe to your local files.
No more manual copy-pasting. No more "Wait, let me find that screenshot of the error."
I asked the AI: "Find the middleware logic that handles the JWT expired state and tell me why it’s failing."
Because Trayce was running in the background, the AI didn't ask for context. It just knew. It pulled the exact line from a file we hadn't even opened yet. Then I typed "find that whiteboard sketch of the auth flow"—and it surfaced a random PNG from three weeks ago instantly.
Took me maybe 5 minutes to get it running. Everything stays local, so I didn't have to worry about our keys or source code leaking to some random cloud vector DB.
Came in after lunch and the lead dev asked me to install it on his machine too. Said it saved him from having to "re-explain his life" to the AI every hour.
The gap between people just "chatting" with AI and people actually using it as a local intelligence layer is where all the speed is. Easiest workflow win of my life, honestly.
Tools I used:
- Trayce — local-first personal intelligence layer. Just indexes everything on your machine and stays out of the way. (usetrayce.com)
- MCP — for the actual context injection so I don't have to copy-paste like a caveman.
- Claude 3.5 Sonnet — still the goat for the actual logic.