# MinTok Autonomous Inference Compiler
Title: MinTok | Autonomous Coding Agents at 1/6th the Inference Cost
Canonical URL: https://mintok.adstim.net/
Documentation: https://mintok.adstim.net/docs/
Benchmarks: https://mintok.adstim.net/benchmark/
Pricing: https://mintok.adstim.net/pricing/
Features: https://mintok.adstim.net/features/
Calculator: https://mintok.adstim.net/tools/token-savings-calculator/
Keywords Model: https://mintok.adstim.net/models/mintok-1-keywords/
Keyword EV Calculator: https://mintok.adstim.net/tools/keyword-ev-calculator/
Title Optimizer: https://mintok.adstim.net/tools/title-optimizer/
Programmatic SEO: https://mintok.adstim.net/solutions/programmatic-seo/

## Overview
MinTok is an autonomous coding inference compiler that optimizes code context before inference. By performing Abstract Syntax Tree (AST) symbol slicing, unreferenced dependency pruning, and observation virtualization, MinTok eliminates up to 84.7% of avoidable context tokens before they reach frontier models.

## Models & Coding Arena Scores
- MinTok 1 Max (mintok-1-max / MinTok-Max alias): 1658 Elo Coding Score | Powered by compiled qwen-3.8-max-0902 | $0.42/M prompt, $1.32/M completion | Frontier tier
- MinTok 1 Pro (mintok-1-pro / MinTok-Pro alias): 1638 Elo Coding Score | Powered by MinTok Context Virtualizer v2 | $0.085/M prompt, $0.24/M completion | Standard tier
- MinTok 1 Flash (mintok-1-flash / MinTok-Flash alias): 1615 Elo Coding Score | Powered by MinTok Sliced IR v2 | $0.05/M prompt, $0.15/M completion | Agile tier
- MinTok 1 Keywords (mintok-1-keywords): LightGBM LambdaRank v2 GBDT (NDCG@10 = 0.9989) | Predictive SEO, Expected Value (eV), Volume, CPC, and SERP difficulty modeling

## Benchmark Performance (SWE-bench / SWE-Holdout-150)
- Solve Rate: 95.3% (143 / 150 tasks solved)
- Average Task Cost: $0.0008 (Flash) to $0.0043 (Max)
- Token Avoidance Ratio: 4.2x (76.2% average context reduction)

## Developer Tools & Installation
- cURL: curl -fsSL https://mintok.adstim.net/install | bash
- npm: npm install -g mintok
- Homebrew: brew install https://mintok.adstim.net/formula/mintok.rb
- Desktop Studio: Available for macOS (Apple Silicon / Intel), Windows, and Linux.

## API Compatibility
- OpenAI-compatible chat completions at https://mintok.adstim.net/v1/chat/completions
- Direct predictive SEO scoring at https://mintok.adstim.net/v1/keywords/score
- Bearer token authentication with mtk_live_ keys
- Real-time Server-Sent Events (SSE) streaming with custom x-mintok- telemetry headers

# MinTok Inference Compiler Technical Architecture Specification

## 4-Stage Context Optimization Pipeline
1. AST Pruning & Subtree Elimination: Language-specific parsers extract caller trees, eliminating unreferenced imports and dead code.
2. Dynamic Slicing: Only variables, classes, and call sites affected by current modifications are retained in context.
3. Observation Virtualization: Shell outputs, compiler errors, and command trajectories are compressed into compact semantic diffs.
4. Prefix-Optimized KV Caching: Static system instructions and repository topologies are isolated at stable prompt prefixes for zero-latency hit rates.

## MinTok CLI Workflow Commands
- /plan <task>: Hierarchical phase breakdown and AST symbol extraction
- /goal <task>: Autonomous agent loop with real-time verification gate
- /side <branch>: Isolated git worktree session
- /review [ref]: Git diff summary and verification verdict
- /status: Active IR phase, symbols, and token budget spent
- /compact: Immediate context eviction and observation virtualization
- /models: List available models and latency profiles
- /doctor: Diagnostic verification of git, python, and backend health
- /cost: Session token spend and net customer savings calculation

## Security & Architecture
- Edge Infrastructure: Cloudflare Workers running inside V8 WebAssembly isolates across 300+ edge PoPs
- Zero Source Code Retention: User code is never persisted to disk or used for training
- Database: Cloudflare D1 serverless relational storage with tenant organization isolation