An MCP plugin that watches every AI output for token waste — fires warnings, errors, and alerts before your context burns.
Detects logs, stack traces, ANSI codes, and repetitive history eating your context window.
Four thresholds — info, warning, error, alert — configurable per project via .token-saver.json.
Works with Claude, GPT, Gemini, or any model output. Any MCP client. Zero lock-in.
One npx -y token-saver-mcp and you're live. No build step needed.
| Model | claude-sonnet-4-6 |
| Tokens used | ~460,000 input · ~72,000 output ~532k total |
| Session duration | ~3 hours (2 sessions, context compacted once) |
| Tech stack | TypeScript · Node.js · MCP SDK · GitHub Actions · semantic-release · Docker · GitHub Pages |
| What was achieved | Designed, built, and shipped token-saver MCP v1.0.0 — full plugin with alert engine, session stats, suppress logic, website, CI/CD, npm publish, and Docker Hub release. |
| Estimated cost | ~$2.46 USD ($3/MTok in · $15/MTok out) |