See where your AI coding money goes.
Self-hosted dashboard that reads your Claude Code, Claude Desktop / Cowork, and Codex CLI session files from disk and shows you — in plain charts — how much you actually spent, on what, and when. No signup. No cloud. Your data never leaves your machine.
If you use AI coding agents every day, you probably have no clue how much that actually costs you.
- Which model burns the most tokens — Opus or Sonnet?
- Which of your projects eats your AI budget?
- Do you spend more on debugging or writing new code?
- Is prompt caching actually saving you anything?
- When do you work, really?
The provider billing dashboards don't answer these. Third Eye does, at per-call granularity, across every session you ever ran — including the ephemeral ones from Cowork.
- Cost breakdowns by model, project, activity, git branch — any date range, any aggregation (day / week / month)
- Per-project drill-down — click any project, get the full story: files you edit most, tools you invoke, subagents you spawn, skills you trigger, MCP servers you lean on
- Activity heatmap — 7×24 grid showing when you actually work on each project
- Claude Code version tracking — see which CLI versions touched each project, distribution by cost / calls / tokens
- Shareable project URLs — stable UUID in the hash, bookmark it, send it
- Everything is local — SQLite file on your disk, you control it and can share it as-is
- 5 languages · Light / Dark / System theme · timezone-aware · mobile-friendly
git clone https://github.com/fien-atone/third-eye
cd third-eye
docker compose up -d --buildOpen http://localhost:4317. The container mounts your ~/.claude and ~/.codex
read-only, re-ingests every 15 minutes, and survives reboots.
git clone https://github.com/fien-atone/third-eye
cd third-eye
npm install
npm startOpen http://localhost:4317. To auto-refresh hourly:
npm run schedule:installIf you'd rather just tell your AI assistant, paste this prompt into Claude Code, Cursor, ChatGPT, or any other coding AI:
I want to install Third Eye, a self-hosted dashboard for AI coding spend. Repo: https://github.com/fien-atone/third-eye
Please do this for me:
- Clone it somewhere sensible under my home directory.
- Pick the best install method for my system — Docker if I have it running (preferred), otherwise Node 20+ via
npm install && npm start.- Start it and verify
http://localhost:4317responds.- Set up hourly auto-ingest so data stays fresh:
- Docker: nothing to do, it's on by default.
- Node: run
npm run schedule:install.- Open http://localhost:4317 in my default browser.
Stop and ask me before making any ambiguous decision (e.g., exposing beyond localhost, picking a non-default port). Show me the URL at the end.
The AI will handle OS quirks, missing tools, and whether you'd rather have Docker or a plain process.
Your session files never leave your machine. Third Eye only reads ~/.claude
and ~/.codex as-is. The server binds to localhost by default. The SQLite
DB lives on your disk — you choose whether and how to share it.
Drilled-in project labels include the first user message of each ephemeral
Cowork session (that's how we turn wizardly-charming-thompson into something
readable). If you share the .db file, recipients can read those prompts. If
that's sensitive, share the code and let each person build their own DB
from their own sessions.
Coming soon — this dashboard is designed for your real data, so the most honest screenshots are the ones you make yourself once it's running.
- DOCS.md — tech stack, API reference, full ingest / scheduler docs, Windows specifics, cost calculation math, timezone handling, dev-mode setup, project layout, data-sharing how-to.
MIT — see LICENSE. © 2026 Ivan Shumov.
The session parser is adapted from CodeBurn by AgentSeal (MIT) — huge thanks for solving the hardest part of the job. Pricing data from LiteLLM. Full attributions in THIRD_PARTY_NOTICES.md.