#105105th of 128 in MCP servers for developers
token-savior
MCP server for symbol-level code navigation, persistent memory and Bash output trimming
- Stars
- 1.2k
- License
- MIT
- Last commit
- Oct 2026
- Last release
- Oct 2026
- Language
- Python
619 on 22 Apr 20261,217 on 11 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Token Savior is a Python MCP server that indexes a codebase by symbol (functions, classes, imports, call graph) so agents fetch pointers instead of whole files. It adds a SQLite WAL + FTS5 memory store re-injected across sessions, plus opt-in Bash hooks: 34 PostToolUse output compactors and a 10-rule PreToolUse command rewriter. The README's 97.9% tsbench score is self-reported and unverifiable; a later 8-task re-measure showed -45% cost.
Who it is for: Claude Code and MCP users cutting token spend on large repos
Strengths
- Symbol index with call graph and structural editing that keeps the index in sync
- Persistent memory in SQLite WAL + FTS5, optional vector search, re-injected next session
- `ts init` merges hooks for Claude, Cursor, Gemini, Codex and OpenClaw with diff and backup
- README openly withdraws a flawed benchmark and reports a re-measure with tool-use checks
Weaknesses
- Headline tsbench 97.9% cannot be reproduced; benchmark repo is not public
- Re-measure on 8 tasks: -45% cost but +40% wall time from indexing
- PostToolUse compactors do not shrink the current turn; only the rewriter does
- Needs `alwaysLoad` on Claude Code, otherwise the model rarely calls its tools
What it needs
- no GPU
- Needs Python 3.11+, SQLite, MCP-compatible client
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Also in MCP servers for developers
See all 128| Rank | Project | Score |
|---|---|---|
| 1 | gitnexusCode knowledge graph exposed to coding agents through MCP tools | 97 out of 100 |
| 2 | codegraphLocal code knowledge graph that coding agents query over MCP | 95 out of 100 |
| 3 | playwright-mcpMCP server that lets LLMs drive a browser through Playwright accessibility snapshots | 93 out of 100 |
| #4 | code-review-graphLocal Tree-sitter code graph that gives coding assistants review context over MCP | 92 out of 100 |
| #5 | fastmcpPython framework for building MCP servers, clients, and interactive apps | 92 out of 100 |