---

DeepTracer

Agents forget each other. Claude’s session does not see Codex. Cursor does not see last week’s Claude run. DeepTracer keeps those logs on this computer, compresses tool calls into meaning, and is the store you query next.

It is not a JSON paste box. You register desktop agents. A local Next.js server reads session files here:

Register agents  →  archive sessions  →  meaning graph  →  recall  →  receipt DAG

A hosted Vercel/Netlify site cannot open ~/.claude, ~/.codex, or ~/.cursor. That is the product, not a limitation to work around.

flowchart TB
  O[Orchestrator]
  O --> R[Research]
  O --> B[Browser]
  R --> S[Vendor A pricing]
  B --> T[Pricing table]

Nodes are what a step obtained — a source, a file, a page — not Search ×4. The graph is the compressed memory. Open a node when you need the original.

Who it reads

Agent	Desktop app	Logs
Claude	Claude	~/.claude/projects, ~/.claude/transcripts
GPT	ChatGPT / Codex	~/.codex/sessions
Cursor	Cursor	~/.cursor/projects/*/agent-transcripts

Only installed desktop apps are listed. Credentials, auth files, and SQLite chat DBs are ignored.

Quick start

git clone https://github.com/acradin/trainthon.git
cd trainthon/deeptracer
npm install
npm run dev

Open http://localhost:3000. Register the agents on this machine. Runs appear without uploading JSON.

Optional deeptracer/.env.local (do not commit):

OPENAI_API_KEY=
OPENAI_MODEL=gpt-5.6-luna

Review intent and meaning-graph compression use OpenAI when a key is present. Without it, traces still scan and open. Default storage is ~/.deeptracer.

Screens

Path	Role
/	Product landing (Recall-first). Registered visits go to /recall
/onboard	Register desktop agents
/recall	Ask what this machine already found
/dashboard	Runs, grouped by project
/trace/[id]	Meaning graph (receipt), Inspector, Timeline
/agents	Status + Scan now
/analyze	Example traces and intent review
/import, /setup	Manual JSON / OTLP fallbacks

Privacy

• Logs stay on the machine running npm run dev.
• The scanner reads session transcripts only.
• Nothing is uploaded unless you add your own OpenAI/Supabase keys.

Docs

App code lives in deeptracer/.

• Product requirements
• Design system
• DAG guidelines
• App README

Trainthon

Built for Cursor Korea × Yonsei University Startup Support Group Trainthon.