The latest articles on DEV Community by Yeskendir Salgara (@sorunokoe). https://dev.to/sorunokoe https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3886533%2F42650408-d258-413f-b673-16c1408f2aab.jpeg https://dev.to/sorunokoe en Yeskendir Salgara Sat, 18 Apr 2026 20:11:53 +0000 https://dev.to/sorunokoe/i-built-a-token-observatory-for-claude-code-because-i-was-tired-of-flying-blind-2n0d https://dev.to/sorunokoe/i-built-a-token-observatory-for-claude-code-because-i-was-tired-of-flying-blind-2n0d <h2> The problem </h2> <p>An hour into a Claude Code session, I got: <strong>"Context window full."</strong></p> <p>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.</p> <p>So I built <strong>Scopeon</strong> - a real-time observability tool that gives you X-ray vision into every token your AI coding agent consumes.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F6bfp7jpbt047otwydi51.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F6bfp7jpbt047otwydi51.png" alt=" " width="800" height="511"></a></p> <h2> Key features </h2> <ul> <li> <strong>Turn-by-turn token breakdown</strong> — see every component, not just totals</li> <li> <strong>Prompt cache ROI</strong> — hit rate gauge + dollars saved vs. uncached</li> <li> <strong>Context countdown</strong> — "~12 turns remaining" before hitting the wall</li> <li> <strong>CI cost gate</strong> — <code>scopeon ci report --fail-on-cost-delta 50</code> fails PRs on AI cost spikes</li> <li> <strong>14 MCP tools</strong> — Claude Code can query its own token stats without polling</li> <li> <strong>Browser dashboard</strong> — live WebSocket charts, zero npm</li> <li> <strong>Fully local</strong> — no cloud, no accounts, SQLite on your machine</li> </ul> <h2> Works with </h2> <p>Claude Code, GitHub Copilot CLI, Aider, Cursor, Gemini CLI, Ollama</p> <h2> Install </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code> curl <span class="nt">-fsSL</span> https://raw.githubusercontent.com/sorunokoe/Scopeon/main/install.sh | sh scopeon onboard </code></pre> </div> <p>The architecture (for the Rust folks)</p> <p>Five crates:</p> <ul> <li>scopeon-core — SQLite schema + pricing engine</li> <li>scopeon-collector — file watchers per provider</li> <li>scopeon-mcp — MCP server (14 tools over stdio)</li> <li>scopeon-tui — ratatui dashboard</li> <li>scopeon-metrics — aggregation + health scoring</li> </ul> <p>Adding a new AI provider takes ~50 lines of Rust.</p> <p>→ GitHub: <a href="https://github.com/sorunokoe/Scopeon" rel="noopener noreferrer">https://github.com/sorunokoe/Scopeon</a></p> rust ai claudecode opensource