Generates a .claudeignore file that excludes Nuxt build artifacts and caches, reducing token usage in Claude sessions.
Generates a .claudeignore file that excludes Storybook build outputs and caches, reducing token usage in Claude sessions.
Generates a .claudeignore file that excludes SvelteKit build artifacts and caches, reducing token usage in Claude sessions.
Generates a .claudeignore file that excludes Terraform artifact files, reducing token usage in Claude sessions.
Generates a .claudeignore file that excludes Turbo (Turborepo) cache directories, reducing token usage in Claude sessions.
Generates a .claudeignore file that excludes Vercel build outputs and caches, reducing token usage in Claude sessions.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
An MCP server plugin for Claude Code that automatically reduces token usage across your sessions. All optimizations are purely algorithmic — no extra API calls, no added cost.
Tool | What it does |
| Strips comments ( |
| Reads only relevant sections of a file. Structure-aware across JS/TS, Python, Go, Rust, Java & C#: returns the complete enclosing function/class/interface around keyword matches, with a configurable fallback window. Rejects binary files |
| Truncates long command output / logs to a token budget. Preserves error/failure lines anywhere in the output, keeps head + tail, collapses duplicate lines |
| Compacts a unified git diff: keeps file headers, hunk headers & changed lines; strips context lines and index/mode noise. Renames and binary files are annotated |
| Counts token usage for any text (cl100k_base encoding) |
| Generates a |
| Rewrites verbose prompts to be concise (~40 filler-phrase rules). Fenced code blocks and inline code are passed through untouched |
All tools return plain text with a compact stats footer — results are deliberately not JSON-wrapped, since JSON escaping of newlines and quotes would inflate the very token count this server exists to reduce.
Note on token counts: the server uses the
cl100k_baseencoding (via tiktoken), which is OpenAI's tokenizer. Claude's tokenizer differs, so all counts are approximations — typically within ~10–20% of Claude's actual usage. Relative savings percentages are unaffected.
Related MCP server: Cozempic
Measured against real code fixtures and realistic prompt inputs. See benchmark/BENCHMARK.md for full methodology.
Tool | Avg token reduction | Best case |
| 31% | 53% on JS with JSDoc |
| 44%* | 81% extracting one function from a module |
| 76% | 84% on long build output |
| 50% | 53% on a multi-file diff with renames |
| 28% | 52% on heavily padded prompts |
| accuracy tool — no reduction metric | — |
| structural correctness tool — no reduction metric | — |
* the smart_read_file average includes tiny synthetic fixtures used as multi-language correctness tests; on realistic files it ranges 38–81%.
Run the benchmark yourself:
npm run benchmarkgit clone <your-repo-url> token-saver-mcp
cd token-saver-mcp
npm installclaude mcp add token-saver -- node /absolute/path/to/token-saver-mcp/src/index.jsOr manually edit ~/.claude.json under mcpServers:
{
"mcpServers": {
"token-saver": {
"command": "node",
"args": ["/absolute/path/to/token-saver-mcp/src/index.js"]
}
}
}claude mcp listYou should see token-saver listed as connected.
Use smart_read_file on src/api/routes.js, focus on "authentication" and "middleware"Generate a .claudeignore for my project at /home/user/myapp and write it to diskCount tokens in this output: [paste output]Compress this before sending: [paste code]Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Exact Claude API cost calc with real cache economics, plus a tiktoken-misuse scanner.
Persistent memory for Claude Code and Cursor. Stop re-explaining your project every session.
Cuts coding agent token burn up to 72% continuously via contradiction-free memory & verification.
Persistent context for Claude. Your AI always knows your projects and next actions across sessions.