AI agents interact with external services through two dominant paradigms today: CLIs which were originally built for humans, and structured tool protocols like MCP. Both impose significant overhead.
AXI is a new paradigm - agent-native CLI tools built from 10 design principles that treat token budget as a first-class constraint.
Results
Browser Benchmark
Evaluated across 490 runs (14 tasks × 7 conditions × 5 repeats) using Claude Sonnet 4.6:
| Condition | Success | Avg Cost | Avg Duration | Avg Turns |
|---|
| chrome-devtools-axi | 100% | $0.074 | 21.5s | 4.5 |
| dev-browser | 99% | $0.078 | 28.6s | 4.9 |
| agent-browser | 99% | $0.088 | 24.6s | 4.8 |
| chrome-devtools-mcp-compressed | 100% | $0.091 | 29.7s | 7.6 |
| chrome-devtools-mcp-search | 99% | $0.096 | 29.4s | 7.5 |
| chrome-devtools-mcp | 99% | $0.101 | 26.0s | 6.2 |
| chrome-devtools-mcp-code | 100% | $0.120 | 36.2s | 6.4 |
GitHub Benchmark
Evaluated across 425 runs (17 tasks × 5 conditions × 5 repeats) using Claude Sonnet 4.6:
| Condition | Success | Avg Cost | Avg Duration | Avg Turns |
|---|
| gh-axi | 100% | $0.050 | 15.7s | 3 |
| gh (CLI) | 86% | $0.054 | 17.4s | 3 |
| GitHub MCP | 87% | $0.148 | 34.2s | 6 |
| GitHub MCP + ToolSearch | 82% | $0.147 | 41.1s | 8 |
| MCP + Code Mode | 84% | $0.101 | 43.4s | 7 |
Claude Sonnet 4.6 is the model the published runs used, not a limit of the
harness. Both benchmarks take --model, which is passed straight through to
the agent CLI, so any Claude model works - for example
--model claude-opus-4-6. Results for other models are simply not published
here, and they may respond differently to output format.
Quick Start
Reference AXI implementations:
See the full AXI Catalog below for all official and community AXIs.
npm install -g gh-axi
npm install -g chrome-devtools-axi
Add to your CLAUDE.md or AGENTS.md:
Use `gh-axi` for GitHub and `chrome-devtools-axi` for browser automation.
The 10 Principles
These principles define what makes a CLI tool "an AXI".
The table below is generated from principles.yaml; the full specification of each principle lives in the AXI skill.
| # | Principle | Summary |
|---|
| 1 | Token-efficient output | Use TOON format for ~40% token savings over JSON |
| 2 | Minimal default schemas | 3–4 fields per list item, not 10+ |
| 3 | Content truncation | Truncate large text with size hints and a --full escape hatch |
| 4 | Pre-computed aggregates | Include aggregated counts and statuses that eliminate round trips |
| 5 | Definitive empty states | Explicit "0 results" rather than ambiguous empty output |
| 6 | Structured errors & exit codes | Idempotent mutations, structured errors, no interactive prompts, fail loud on unknown flags |
| 7 | Ambient context | Install opt-in session integrations first, then offer an on-demand skill |
| 8 | Content first | Running with no arguments shows live data, not help text |
| 9 | Contextual disclosure | Include next-step suggestions after each output |
| 10 | Consistent way to get help | Concise per-subcommand reference when agents need it |
AXI Catalog
The catalog tables below are generated from catalog.yaml - see CONTRIBUTING.md to add your AXI.
Official
Reference implementations maintained by the AXI project, validating the principles across different domains:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
