TokMan πΈ
World's Most Advanced Token Reduction System β 14-layer research-based compression pipeline achieving 95-99% token reduction.
TokMan intercepts CLI commands, applies 14 research-backed compression layers, and tracks token savings. Built on 50+ papers from top institutions (Mila, Microsoft, Stanford, Berkeley, NeurIPS 2023-2025).
| Input Size |
Original |
Final |
Reduction |
| Small (100 lines) |
982 tokens |
44 tokens |
95.5% |
| Medium (1000 lines) |
9,737 tokens |
52 tokens |
99.5% |
| Large (5000 lines) |
49,437 tokens |
63 tokens |
99.9% |
| Up to 2M tokens |
Full context support |
Streaming processing |
Memory optimized |
Features
- π§ 14-Layer Compression Pipeline β Research-based token reduction (95-99%)
- π§ Git Command Wrappers β Filtered
status, diff, log, add, commit, push, pull, and more
- π LS Handler β Hide noise directories (.git, node_modules, target, etc.)
- π³ Infrastructure Wrappers β Docker, kubectl, AWS CLI with filtered output
- π¦ Package Managers β npm, pnpm, pip, cargo with compact output
- π§ͺ Test Runners β Go, pytest, vitest, jest, npm test, playwright with aggregated results
- π¨ Build Tools β Go, cargo, next.js with error-only output
- π Token Tracking β SQLite-based metrics on tokens saved
- π Shell Integration β Automatic command rewriting via shell hooks
- π Integrity Verification β SHA-256 hook verification for security
- π° Economics Analysis β Compare spending vs savings with quota estimates
- πΎ Tee on Failure β Auto-saves full output when commands fail for debugging
14-Layer Compression Pipeline
TokMan implements the world's most advanced token reduction system based on 50+ research papers:
| Layer |
Name |
Research |
Compression |
| 1 |
Entropy Filtering |
Selective Context (Mila 2023) |
2-3x |
| 2 |
Perplexity Pruning |
LLMLingua (Microsoft 2023) |
20x |
| 3 |
Goal-Driven Selection |
SWE-Pruner (Shanghai Jiao Tong 2025) |
14.8x |
| 4 |
AST Preservation |
LongCodeZip (NUS 2025) |
4-8x |
| 5 |
Contrastive Ranking |
LongLLMLingua (Microsoft 2024) |
4-10x |
| 6 |
N-gram Abbreviation |
CompactPrompt (2025) |
2.5x |
| 7 |
Evaluator Heads |
EHPC (Tsinghua/Huawei 2025) |
5-7x |
| 8 |
Gist Compression |
Stanford/Berkeley (2023) |
20x+ |
| 9 |
Hierarchical Summary |
AutoCompressor (Princeton/MIT 2023) |
Extreme |
| 10 |
Budget Enforcement |
Industry standard |
Guaranteed |
| 11 |
Compaction |
MemGPT (UC Berkeley 2023) |
98%+ |
| 12 |
Attribution Filter |
ProCut (LinkedIn 2025) |
78% |
| 13 |
H2O Filter |
Heavy-Hitter Oracle (NeurIPS 2023) |
30x+ |
| 14 |
Attention Sink |
StreamingLLM (2023) |
Infinite stability |
See docs/LAYERS.md for detailed documentation of each layer.
Installation
# Clone and build
git clone https://github.com/GrayCodeAI/tokman.git
cd tokman
go build -o tokman ./cmd/tokman
# Install to PATH (optional)
sudo mv tokman /usr/local/bin/
Docker
# Pull from GitHub Container Registry (once published)
docker pull ghcr.io/graycodeai/tokman:latest
# Or build locally
docker build -f docker/Dockerfile -t tokman:latest .
# Run with dashboard
docker run -d -p 8080:8080 -v tokman-data:/home/tokman/.local/share/tokman tokman:latest
# Run with docker-compose
cd docker && docker-compose up -d
Homebrew (macOS/Linux)
brew install GrayCodeAI/tap/tokman
Quick Start
# Initialize TokMan (install shell hook)
tokman init
# Check token savings
tokman status
# View comprehensive savings analysis
tokman gain
# Use wrapped commands
tokman git status
tokman ls
tokman test ./...
Commands
Core Commands
| Command |
Description |
tokman init |
Initialize TokMan and install shell hook |
tokman status |
Quick token savings summary |
tokman report |
Detailed usage analytics |
tokman gain |
Comprehensive savings with graphs, history, quota |
tokman config |
Show or create configuration file |
tokman verify |
Verify hook integrity (SHA-256 check) |
tokman economics |
Show spending vs savings analysis |
Git Wrappers
| Command |
Filter Applied |
tokman git status |
Porcelain parsing, emoji formatting |
tokman git diff |
Stats summary, compact hunks |
tokman git log |
Oneline format, smart limits |
tokman git add |
Compact "ok β" output |
tokman git commit |
Show hash on success |
tokman git push |
Show branch on success |
tokman git pull |
Show stats summary |
tokman git branch |
Compact listing |
tokman git stash |
Compact list/apply/drop |
tokman git show |
Commit summary + compact diff |
tokman git fetch |
Show new refs count |
tokman git worktree |
Compact listing |
Infrastructure Wrappers
| Command |
Description |
tokman docker |
Docker CLI with filtered output |
tokman kubectl |
Kubernetes CLI with filtered output |
tokman aws |
AWS CLI with filtered output |
tokman gh |
GitHub CLI with token-optimized output |
tokman gh run list |
GitHub Actions workflow runs (compact) |
tokman gh release list |
GitHub releases (compact) |
tokman gh api |
GitHub API with JSON structure output |
tokman gt |
Graphite stacked PR commands |
Test Runners
| Command |
Description |
tokman go test |
Go tests with aggregated results |
tokman go build |
Go build with error-only output |
tokman go vet |
Go vet with compact output |
tokman cargo test |
Cargo tests with compact output |
tokman cargo build |
Cargo build with error-only output |
tokman cargo clippy |
Rust linter with compact output |
tokman pytest |
Python tests with compact output |
tokman ruff |
Python linter/formatter compact |
tokman mypy |
Python type checker compact |
tokman vitest |
Vitest with compact output |
tokman jest |
Jest with compact output (90% reduction) |
tokman npm test |
npm test with compact output (90% reduction) |
tokman playwright |
Playwright E2E tests compact |
Package Managers
| Command |
Description |
tokman npm |
npm run with filtered output |
tokman pnpm |
pnpm with ultra-compact output |
tokman npx |
npx with intelligent routing |
tokman pip |
pip with compact output |
tokman cargo |
Cargo commands with filtering |
Utilities
| Command |
Description |
tokman ls |
Hide noise dirs, human sizes |
tokman tree |
Compact tree output |
tokman find |
Find files with compact output |
tokman grep |
Compact grep, groups by file |
tokman diff |
Ultra-condensed diff |
tokman json |
Show JSON structure |
tokman env |
Show env vars (sensitive masked) |
tokman deps |
Summarize dependencies |
tokman log |
Filter/deduplicate logs |
tokman wc |
Word/line/byte count compact |
tokman curl |
Auto-JSON detection |
tokman wget |
Download with compact output |
tokman summary |
Heuristic summary of long output |
tokman count |
Count tokens using tiktoken (OpenAI tokenizer) |
Analysis Commands
| Command |
Description |
tokman discover |
Find missed savings in Claude Code history |
tokman learn |
Generate CLI correction rules from errors |
tokman err <cmd> |
Run command, show only errors/warnings |
tokman proxy <cmd> |
Run without filtering (still tracked) |
tokman hook-audit |
Show hook rewrite metrics |
Rewriting
| Command |
Description |
tokman rewrite <cmd> |
Rewrite a command to use TokMan |
tokman rewrite list |
List all registered rewrites |
Shell Integration
Token Counting (tiktoken)
TokMan includes direct token counting using OpenAI's tiktoken tokenizer:
# Count tokens in text
tokman count "Hello, world!"
# Count tokens from stdin
cat file.txt | tokman count
# Count tokens in a file
tokman count file.go
# Use specific model encoding
tokman count --model gpt-4o "Hello, world!"
tokman count --model claude-3-sonnet "Hello, world!"
# Compare heuristic vs actual count
tokman count --compare "Your text here"
# Count multiple files
tokman count --files *.go
Supported encodings:
cl100k_base β GPT-4, GPT-3.5-turbo, Claude, text-embedding-ada-002
o200k_base β GPT-4o, GPT-4o-mini
p50k_base β GPT-3 (davinci, curie, babbage, ada)
Add to your .bashrc or .zshrc:
source /path/to/tokman/hooks/tokman-rewrite.sh
This enables:
- Automatic command rewriting for all supported commands
ts β alias for tokman status
tr β alias for tokman rewrite
tokman_install_hook β install hook to shell config
tokman_status β show integration status
Shell Completions
Enable autocompletions for your shell:
# Bash
source <(tokman completion bash)
# Zsh
source <(tokman completion zsh)
# Fish
tokman completion fish | source
Or save the completion files from completions/ directory.
Custom Filter Plugins
Create JSON-based filter plugins in ~/.config/tokman/plugins/:
{
"name": "hide-npm-warnings",
"description": "Hide npm deprecation warnings",
"enabled": true,
"patterns": ["npm WARN deprecated"],
"mode": "hide"
}
# Plugin management
tokman plugin list # List loaded plugins
tokman plugin create myfilter # Create new plugin template
tokman plugin enable myfilter # Enable a plugin
tokman plugin disable myfilter # Disable a plugin
tokman plugin examples # Generate example plugins
Web Dashboard
Launch an interactive dashboard to visualize token savings:
# Start dashboard on default port (8080)
tokman dashboard
# Custom port
tokman dashboard --port 3000
# Open browser automatically
tokman dashboard --open
Features:
- Real-time token savings charts
- Daily/weekly/monthly breakdowns
- Command-level analytics
- Cost tracking with Claude API rates
- RESTful API endpoints (
/api/stats, /api/daily, /api/commands)
CI/CD Integration
Include TokMan in your CI pipelines for automated reporting:
GitHub Actions
# .github/workflows/tokman.yml
name: Token Savings Report
on: [workflow_run]
jobs:
report:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-go@v5
with: { go-version: '1.21' }
- run: go install github.com/GrayCodeAI/tokman/cmd/tokman@latest
- run: tokman summary --format markdown >> $GITHUB_STEP_SUMMARY
See templates/ for complete GitHub Actions and GitLab CI examples.
Configuration
Config file: ~/.config/tokman/config.toml
[tracking]
enabled = true
telemetry = false
[filter]
mode = "minimal" # "minimal" or "aggressive"
noise_dirs = [
".git", "node_modules", "target",
"__pycache__", ".venv", "vendor",
".idea", ".vscode", "dist", "build"
]
[hooks]
excluded_commands = []
Gain Command Features
The tokman gain command provides comprehensive savings analysis:
# Basic summary
tokman gain
# Show ASCII graph of daily savings
tokman gain --graph
# Show recent command history
tokman gain --history
# Show quota analysis (subscription tier estimate)
tokman gain --quota --tier pro
# Time breakdowns
tokman gain --daily
tokman gain --weekly
tokman gain --monthly
tokman gain --all
# Export data
tokman gain --format json
tokman gain --format csv
# Filter to current project
tokman gain --project
Security: Integrity Verification
TokMan protects against hook tampering:
# Verify hook integrity
tokman verify
# Runtime checks are automatic for operational commands
# Hooks are verified via SHA-256 hash stored during init
Architecture
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β TokMan CLI β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Command Handlers (internal/commands/) β
β βββ git.go - Git wrappers β
β βββ docker.go - Docker filtering β
β βββ go.go - Go build/test β
β βββ cargo.go - Rust commands β
β βββ npm.go - npm/pnpm/npx β
β βββ pytest.go - Python tests β
β βββ gain.go - Savings analysis β
β βββ economics.go - Cost analysis β
β βββ ... - 40+ command handlers β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β 14-Layer Compression Pipeline (internal/filter/) β
β βββ Layer 1-9: Research compression (95%+) β
β βββ Layer 10: Budget enforcement β
β βββ Layer 11: Compaction (MemGPT-style) β
β βββ Layer 12: Attribution (ProCut-style) β
β βββ Layer 13: H2O (Heavy-Hitter Oracle) β
β βββ Layer 14: Attention Sink (StreamingLLM) β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Core Engine (internal/) β
β βββ tracking/ - SQLite token tracking β
β βββ config/ - TOML config loader β
β βββ integrity/ - SHA-256 hook verification β
β βββ economics/ - Cost analysis engine β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Shell Integration (hooks/) β
β βββ tokman-rewrite.sh - Bash/Zsh hook β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
Development
# Run tests
go test ./...
# Build
go build -o tokman ./cmd/tokman
# Run with verbose output
./tokman -v git status
Token Estimation
TokMan uses a simple heuristic for token estimation:
tokens = ceil(text.length / 4.0)
This provides a reasonable approximation for tracking savings without requiring an actual tokenizer.
Data Storage
Following XDG Base Directory Specification:
| Resource |
Path |
| Config |
~/.config/tokman/config.toml |
| Database |
~/.config/tokman/tracking.db |
| Logs |
~/.config/tokman/tokman.log |
| Hook |
~/.claude/hooks/tokman-rewrite.sh |
| Hook Hash |
~/.claude/hooks/tokman-rewrite.sh.sha256 |
Override with environment variables:
XDG_CONFIG_HOME
TOKMAN_DATABASE_PATH
Roadmap
-
Git command wrappers
-
Token tracking database
-
Shell integration hooks
-
Integrity verification
-
Economics analysis
-
Python tool wrappers (pytest, ruff, mypy)
-
Rust/Cargo test aggregation
-
Windows support
-
Custom filter plugins
-
Web dashboard for analytics
-
Shell completions (bash/zsh/fish)
-
CI/CD integration templates
- LLM API integration (direct token counting)
- Homebrew formula
- Docker image
License
MIT License - see LICENSE for details.
πΈ TokMan β Save tokens, save money, save the context window.