The problem
An hour into a Claude Code session, I got: "Context window full."
I had no idea what burned it. Was it the MCP tools fetching docs? The thinking budget? The growing list of edited files? I was flying blind on a meter that was costing me real money.
So I built Scopeon - a real-time observability tool that gives you X-ray vision into every token your AI coding agent consumes.
Key features
- Turn-by-turn token breakdown — see every component, not just totals
- Prompt cache ROI — hit rate gauge + dollars saved vs. uncached
- Context countdown — "~12 turns remaining" before hitting the wall
-
CI cost gate —
scopeon ci report --fail-on-cost-delta 50fails PRs on AI cost spikes - 14 MCP tools — Claude Code can query its own token stats without polling
- Browser dashboard — live WebSocket charts, zero npm
- Fully local — no cloud, no accounts, SQLite on your machine
Works with
Claude Code, GitHub Copilot CLI, Aider, Cursor, Gemini CLI, Ollama
Install
curl -fsSL https://raw.githubusercontent.com/sorunokoe/Scopeon/main/install.sh | sh
scopeon onboard
The architecture (for the Rust folks)
Five crates:
- scopeon-core — SQLite schema + pricing engine
- scopeon-collector — file watchers per provider
- scopeon-mcp — MCP server (14 tools over stdio)
- scopeon-tui — ratatui dashboard
- scopeon-metrics — aggregation + health scoring
Adding a new AI provider takes ~50 lines of Rust.
→ GitHub: https://github.com/sorunokoe/Scopeon
