Tokenectomy 🕵️♂️ — Autonomous Context Surgery & Secret Shield for AI Agents (M2M MCP Server)
👑 Looking for Auto-Fixer, AST Smart Healer, Anti-Halu, and Time Machine Undo? Get Tokenectomy Pro (Enterprise Edition) →
🔀 Pair with Tokenectomy Git (OSS) for autonomous Git branch, commit, and Pull Request creation!
Tokenectomy is an agent-native Machine-to-Machine (M2M) MCP server built with Rust. Designed specifically as a background sidecar for autonomous coding agents (Claude Desktop, Cursor, Cline, Roo Code, Windsurf, Google Antigravity), it acts as an autonomous sub-cortex: surgically scrubbing 90%+ of internal framework noise (node_modules, site-packages, .cargo/registry) from error logs, auto-redacting sensitive credentials before cloud transmission, and enforcing AST syntax safety—with zero human babysitting.
┌──────────────────┐
Agent Error Dump │ TOKENECTOMY │ Clean Agent Context
(38K tokens) ───► │ 🕵️♂️ OSS (M2M) │ ───► (2K tokens) ───► LLM Brain
│ │
node_modules/ │ 🔍 Smart Filter │ Only YOUR code
site-packages/ │ 🛡️ Redact │ + error message
.cargo/registry/ │ 💾 Cache │ + StackOverflow refs
└──────────────────┘
✨ Key Features
- 🔍 Smart Framework Filter — Strips thousands of lines of noisy internal stack frames (
node_modules,site-packages,.cargo/registry,__pycache__) and keeps only the code you wrote. - 🌐 Stack Overflow Search — Silently queries StackExchange APIs and injects top community solutions into the AI's context.
- ⚡ SHA-256 Response Cache — Identical errors hit local cache (24h TTL). Recurring CI/CD failures cost $0.00 in API calls.
- 🛡️ Secret Redaction — Regex engine strips API keys, AWS secrets, JWTs, and database connection strings before any data leaves your machine (ReDoS-safe, linear-time).
- 🔒 Anti-Hardcode Secret Shield — Automatically detects and blocks AI patches that attempt to hardcode raw API keys, passwords, or credentials into your source code.
- 📐 AST Syntax Validation — In-memory Tree-sitter AST parser ensures AI patches never write broken syntax to your repository.
- 🔒 Path Traversal Protection — MCP file operations are canonicalized and locked to your current working directory.
- 🤖 MCP Server Mode — Full JSON-RPC 2.0 over stdio. Works with Claude Desktop, Cursor, VS Code, Google Antigravity, and any MCP-compatible client.
- 🔌 Multi-Provider — Supports OpenAI, Anthropic, and Ollama (100% offline mode).
📊 Verifiable Real-World Performance & Benchmark (OSS vs Pro)
No buzzwords or artificial benchmarks. Every developer can verify the core functions on their own machine:
🆓 Tokenectomy OSS (Community Edition) — Verifiable Heavy Stress Benchmark
You don't need to buy anything to test this. Clone this repository right now and verify these heavy load benchmarks directly on your hardware:
| Feature Under Test | Tested Heavy Input | Real Measured Outcome | Status |
|---|---|---|---|
| Massive Log Redaction | 25,000 lines (2.41 MB) dump with DB URLs, API keys, JWTs | 65.79 ms (36.7 MB/sec, 379,997 lines/sec). 100% sanitized. | ✅ Verified |
| ReDoS Immunity | 50,000-character malicious backtracking exploit string | 7.11 ms. Linear $O(N)$ evaluation, 100% ReDoS immune. | ✅ Verified |
| High Concurrency Torture | 100 concurrent OS threads hammering redaction & extractor | 100/100 in 265 ms (752.2 ops/sec). Zero race conditions. | ✅ Verified |
| Kernel Memory Footprint | Peak Resident Memory during 25K-line stress test | 14.11 MB VmRSS via Linux /proc/self/status. Zero ballooning. |
✅ Verified |
💡 "Skeptical about these numbers? Don't take our word for it."
We hate marketing fluff and sweet talk as much as you do. You don't need to take our word for it or pay a single cent. Clone this repository, run the benchmark on your own machine, watch your CPU blaze through 25,000 lines of logs in milliseconds, and verify the exact telemetry in your own terminal:
👑 Tokenectomy Pro — Industrial-Grade Heavy Production Torture Benchmark
To prove stability under enterprise workloads, we hammered Tokenectomy Pro with sustained multi-million token traffic, a 10MB Kubernetes crash avalanche, and 250 parallel OS threads on a 12-Core Intel i5-1235U with Arch Linux:
| Stress Vector | Tested Workload | Measured Kernel / CPU Telemetry | Status |
|---|---|---|---|
| Sustained Stream | 100 consecutive microservice crash incidents (2.16 MB) | 604,490 ➔ 6,490 tokens (98.93% reduction) at 41,752 tokens/sec. | ✅ Passed |
| K8s Crash Avalanche | 6,506-line dump combining Spring Boot, PyTorch OOM, & Go | Processed in 514 ms (1.6 MB/sec). User code preserved across 3 languages. | ✅ Passed |
| Monorepo Parallel AST | 100 multi-language files (Rust, TS, Python, Go) parsed simultaneously | 2.68 ms total (0.027 ms / file) = 37,379 files/sec. Zero memory leak. | ✅ Passed |
| Extreme Concurrency | 250 parallel OS threads hammering AST, Redact, & HaluGuard | 250/250 passed in 865 ms (867 ops/sec). Zero deadlock or race condition. | ✅ Passed |
| Memory Footprint (VmRSS) | Full 600K-token & 250-thread torture test | Peak RAM capped at 83.77 MB via Linux /proc/self/status. Zero leak. |
✅ Controlled |
💡 "Skeptical about these enterprise numbers? Don't take our word for it."
Every purchaser of Tokenectomy Pro receives the full, standalone benchmark and torture test suite (tests/heavy_production_torture.rs) bundled in the package. You can runcargo test --release --test heavy_production_torture -- --nocaptureon your own infrastructure to verify every metric before deployment.
👉 Get Tokenectomy Pro on Gumroad ($9) → (Native binaries for Linux, macOS Apple Silicon, and Windows).
📦 Installation
⚡ Install via Cargo (crates.io)
🦀 Build from Source
⚙️ Install via Smithery (for Claude Desktop)
🔌 M2M Agent Setup (1-Minute Integration)
Tokenectomy is architected to run silently between your AI Coding Agent and your repository over JSON-RPC 2.0 stdio. You configure it once, and your agent autonomously invokes Tokenectomy in the background during debugging and refactoring loops—no manual copy-pasting or piping required.
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
Cursor
Add to .cursor/mcp.json in your project root:
Cline / Roo Code / Windsurf / VS Code
Add to your MCP settings (settings.json or cline_mcp_settings.json):
Google Antigravity CLI
🤖 Available M2M MCP Tools
| Tool | Autonomous Agent Role |
|---|---|
get_error_context |
Performs deep log surgery: strips framework noise, redacts secrets, extracts source context and git diff |
search_stack_overflow |
Searches Stack Overflow for a specific error (query is auto-sanitized of secrets) |
apply_code_patch |
Applies a code patch to a file by search-and-replace |
🔀 Companion MCP Server: Tokenectomy Git
Close the autonomous loop from error diagnosis all the way to a published GitHub Pull Request! Pair Tokenectomy with our official companion MCP server: Tokenectomy Git.
Together, they enable your AI coding assistant to:
- Scrub noisy error logs & redact credentials (
tokenectomy) - Generate an accurate fix patch
- Create a branch, commit files, and open a GitHub PR autonomously (
tokenectomy-git)
🛠️ Standalone / Local CLI Mode (Optional)
While Tokenectomy is architected for autonomous machine-to-machine agent operation, it also provides a standalone CLI binary if you want to pipe logs in CI/CD pipelines, local shell scripts, or manual debugging:
Pipe errors directly
|
|
|
Read from a log file
Advanced options
🧠 Configuration
Create ~/.tokenectomy.toml:
= "openai" # openai | anthropic | ollama | mock
= "sk-..."
= "sk-ant-..."
= "http://localhost:11434"
= 10
= 10000
🔒 Security
- No secrets leave your machine. Regex engine redacts API keys, JWTs, AWS credentials, and database URLs before any data is sent to an LLM.
- No path traversal. MCP file operations are canonicalized and locked to CWD.
- No stdin bombs. Input is capped at 10MB (CLI) / 50MB (MCP) via
.take(). - No weak hashing. Cache uses
sha2::Sha256, neverDefaultHasher.
🏗️ Architecture
src/
├── main.rs # Entry point, CLI args, REPL, pipeline orchestration
├── mcp.rs # JSON-RPC 2.0 over stdio MCP server
├── extractor/ # Language-specific context extraction (Rust, Python, JS)
├── provider/ # AI backends (OpenAI, Anthropic, Ollama, Mock)
├── redact.rs # Secret redaction (linear-time regex, ReDoS-safe)
├── cache.rs # SHA-256 response cache (24h TTL, 0700 perms)
├── search.rs # Stack Overflow API integration
└── git.rs # Recent git diff extraction
🧠 Bundled Open Source Agent Skills
Tokenectomy OSS includes 2 native Antigravity & Coding Agent Skills in .agents/skills/ to elevate your AI assistant's engineering discipline:
| Skill | Description | Location |
|---|---|---|
adversary-bug-hunter |
Red-team fuzzer that stress-tests edge cases, catches unhandled unwraps/ReDoS, and hardens code. | .agents/skills/adversary-bug-hunter/ |
spec-first-architect |
Enforces strict Test-Driven Development (TDD) & state machine invariants to eliminate AI hallucinations. | .agents/skills/spec-first-architect/ |
🆓 vs 👑 — OSS vs Pro
| Feature | OSS (Free) | Pro ($9) |
|---|---|---|
| Smart Framework Filter | ✅ | ✅ |
| Stack Overflow Search | ✅ | ✅ |
| SHA-256 Response Cache | ✅ | ✅ |
| Secret Redaction (ReDoS-safe) | ✅ | ✅ |
| Anti-Hardcode Secret Shield | ✅ | ✅ |
| AST Syntax Validation (Tree-sitter) | ✅ | ✅ |
| MCP Server Mode | ✅ | ✅ |
| Multi-Provider (OpenAI, Anthropic, Ollama) | ✅ | ✅ |
| Bundled Agent Skills | 2 Skills | 4 Full Skills Suite |
| Auto-Fixer (AI patch → auto-apply) | ❌ | ✅ |
| AST Smart Healer (auto-repair syntax errors) | ❌ | ✅ |
| Code Integrity Guard (Anti-Halu & Anti-Ngide) | ❌ | ✅ |
| Test Verification Loop (auto-rollback on test fail) | ❌ | ✅ |
| Multi-File Atomic Transactions | ❌ | ✅ |
Time Machine Undo Engine (1-sec revert via tkmy --undo) |
❌ | ✅ |
| True Ectomy Engine (99% token reduction) | ❌ | ✅ |
| DB Inspector (real TCP port probe) | ❌ | ✅ |
| Docker Diagnostics (OOMKilled detection) | ❌ | ✅ |
--benchmark Mode |
❌ | ✅ |
| SRE Incident Commander Skill (OOM 137 triage & 5-Whys RCA) | ❌ | ✅ Included |
| Refactor Sentinel Skill (Zero-regression blast radius refactor) | ❌ | ✅ Included |
📜 License
MIT — see LICENSE for details.
Built with 🦀 Rust for maximum performance, strict memory safety, and uncompromising security.