ArbiterX is intelligent middleware for AI coding assistants (Claude Code, Codex, Cursor, Copilot, Aider). It reduces tokens by 97% via codebase mapping, routes tasks to the right model based on complexity, and scores all generated code 0-100 with a quality gate that checks security, robustness, and efficiency.

How does ArbiterX reduce tokens by 97%?

ArbiterX parses your codebase with tree-sitter, extracts function/class definitions and their relationships, ranks them by relevance using PageRank, and stores them in SQLite. When the AI needs context, it queries the map and gets only relevant signatures (5,000 tokens) instead of reading full files (200,000 tokens).

How do I install ArbiterX?

Install with pip: pip install arbiterx-gate. Then run arbiterx init in your project and arbiterx map to index your codebase. For AI tool integration, install the plugin: /plugin install NeelPrime/arbiterx in Claude Code.

What AI coding tools does ArbiterX work with?

ArbiterX works with Claude Code, Codex CLI, Cursor, GitHub Copilot, Aider, Windsurf, Kiro, Zed, and any tool that reads AGENTS.md.

Yes. ArbiterX is free and open source under the Apache-2.0 license. Zero telemetry. Your code never leaves your machine.

ArbiterX — AI Code Quality Middleware | Smart Routing, 97% Token Reduction, Quality Gate

Three Superpowers

ArbiterX sits between you and your AI tool. Here's what it does.

🗺️

Codebase Map

Parses your code with tree-sitter, ranks symbols with PageRank, and sends only what the AI needs. 200K tokens → 5K tokens. That's 97% savings on every query.

🧠

Smart Router

Classifies tasks by complexity. Renames go to Haiku ($0.25/M). Bug fixes go to Sonnet ($3/M). Architecture goes to Opus ($15/M). Same quality, fraction of cost.

🛡️

Quality Gate

Scores every AI output 0–100. Checks security, robustness, efficiency, types, style, completeness. Rejects code scoring below 70. No more reviewing AI slop.

See It in Action

Three commands. You're running.

arbiterx — terminal

$ arbiterx map

Files indexed: 46

Symbols extracted: 1,364

Build time: 0.25s

$ arbiterx route "fix the auth bypass bug"

Type: DEBUGGING

Complexity: HIGH

→ Model: opus (deep security reasoning)

$ arbiterx gate --file bad_code.py

❌ Score: 35/100 FAILED

• [high] Hardcoded password detected

• [med] File operation without error handling

$ arbiterx gate --file good_code.py

✅ Score: 100/100 PASSED

Quick Start

Install in seconds. Works with your existing AI tool.

$ pip install arbiterx-gate
$ arbiterx init
$ arbiterx map              # Index your codebase (~30s)

# Use with Claude Code:
$ /plugin install NeelPrime/arbiterx

# Use with Codex CLI:
$ codex plugin install NeelPrime/arbiterx

# Use with Cursor, Copilot, Aider:
$ arbiterx export-rules --format cursor
$ arbiterx export-rules --format copilot
$ arbiterx export-rules --format agents   # universal

Works With Everything

Plugin install for Claude Code & Codex. Rules file for everything else.

Claude Code/plugin install

Codex CLIplugin install

Cursor.cursor/rules/

GitHub Copilotinstructions.md

AiderAGENTS.md

Windsurf.windsurf/rules/

Kiro.kiro/steering/

Zed.zed/assistant/

Git Hookspre-commit

Any ToolAGENTS.md

The 10 Engineering Rules

Enforced on every piece of generated code. No exceptions.

01YAGNI — don't build what wasn't asked for

02Handle every error path

03Type all function signatures

04Close every resource (context managers)

05Name all constants — no magic numbers

06No dead code — no commented code

07Single responsibility — max 30 lines

08Fail fast — validate inputs at entry

09Idempotency — safe to retry

10Performance — no unnecessary O(n²)

Frequently Asked Questions

What is ArbiterX?

ArbiterX is intelligent middleware for AI coding assistants. It sits between you and your AI tool (Claude Code, Codex, Cursor, Copilot) and compresses context by 97%, routes tasks to the optimal model, and scores all generated code with a quality gate (0-100).

How does it reduce tokens by 97%?

It parses your codebase with tree-sitter, extracts function and class definitions, ranks them by importance using PageRank, and stores them in SQLite. The AI gets relevant signatures (5K tokens) instead of full files (200K tokens).

Does it slow down my workflow?

No. Map builds in under 1 second. Symbol queries take 0.5ms. The quality gate uses regex-based static analysis — no LLM calls needed for scoring.

Does it work offline?

Yes. The map, router, and quality gate all run locally. Only the LLM adapters need internet. Use Ollama for fully offline AI generation.

Is it free?

Yes. Apache-2.0 license. Zero telemetry. Your code never leaves your machine.

How is this different from a linter?

Linters check syntax. ArbiterX enforces engineering decisions — YAGNI, proper abstractions, security patterns. It also handles context compression (97% fewer tokens) and model routing, which linters don't do.

What languages are supported?

22 languages: Python, TypeScript, JavaScript, Go, Rust, Java, C, C++, C#, Ruby, PHP, Swift, Kotlin, Scala, Elixir, Haskell, OCaml, Lua, Zig, Bash, JSX, and TSX.

What does the quality gate check?

Security (hardcoded secrets, SQL injection), Robustness (error handling, resource cleanup), Efficiency (O(n²) patterns), Type Safety, Style, and Completeness (no TODOs). Code below 70/100 is rejected.