Kun ChenAXI: Agent eXperience Interface
A new standard for agent ergonomics. It beats both MCP and regular CLI in success rate and efficiency.
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 — CLIs produce verbose, metadata-sparse output that wastes token budgets, while MCP tool schemas consume a lot of context tokens upfront.
AXI is a new paradigm — agent-native CLI tools built from 10 design principles that treat token budget as a first-class constraint. An AXI provides token-optimized output formatting, pre-computed aggregate fields, contextual next-step suggestions, and structured error handling. gh-axi is the first AXI — a gh wrapper you can install today and point your agent at via CLAUDE.md or AGENTS.md.
Results
Evaluated across 425 benchmark 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 |
Quick Start
gh-axi is the reference AXI implementation — an ergonomic wrapper around the gh CLI.
$ npm install -g gh-axi
Add to your CLAUDE.md or AGENTS.md:
Use `gh-axi` (replacement for `gh` CLI) for all GitHub operations.
That's it. Your agent now gets structured, token-efficient GitHub output.
The 10 Principles
These principles define what makes a CLI tool "an AXI":
| # | Principle | Summary |
|---|---|---|
| 1 | Token-efficient output | Use TOON format for ~40% token savings over JSON |
| 2 | Content first | Running with no arguments shows live data, not help text |
| 3 | Contextual disclosure | Include next-step suggestions after each output |
| 4 | Provide --help | Concise per-subcommand reference when agents need it |
| 5 | Minimal default schemas | 3–4 fields per list item, not 10 |
| 6 | Pre-computed fields | Include aggregated statuses that eliminate round trips |
| 7 | Content truncation | Truncate large text with size hints and --full escape hatch |
| 8 | Definitive empty states | Explicit "0 results" rather than ambiguous empty output |
| 9 | Error handling | Idempotent mutations, structured errors, no interactive prompts |
| 10 | Output discipline | stdout for data, stderr for debug; clean exit codes |
Build Your Own AXI
Install the AXI skill to get the design guidelines and scaffolding for building an AXI-compliant CLI:
$ npx skills add kunchenguid/axi
This installs the AXI skill — a detailed guide with examples for each principle that your coding agent can reference while building.
gh-axi
gh-axi wraps the gh CLI with token-efficient TOON output, pre-computed fields, contextual suggestions, and structured errors. See the gh-axi README for full usage and command reference.
$ gh-axi pr list --state merged --limit 3
count: 3 of 3549
pull_requests[3]{number,title,state,author}:
51772,"refactor(plugins): route Telegram...",merged,dependabot
51770,"fix(core): handle nil pointer in...",merged,alice
51769,"feat(auth): add OAuth2 PKCE flow",merged,bob
help[1]:
Run `gh-axi pr view <number>` to see full details
Development
gh-axi
cd gh-axi
npm install # Install dependencies
npm run dev # Run in development (tsx)
npm run build # Build (tsc)
npm test # Run tests (vitest)
Requires Node.js >= 20 and the GitHub CLI (gh) installed and authenticated.
Running the Benchmark
The benchmark harness lives in bench/. It runs agent tasks across different interface conditions and grades results with an LLM judge.
cd bench
npm install
# Run a single condition × task
npm run bench -- run --condition axi --task merged_pr_ci_audit --repeat 5 --agent claude
# Run the full matrix (all conditions × all tasks)
npm run bench -- matrix --repeat 5 --agent claude
# Generate summary report from results
npm run bench -- report
Conditions: axi, cli, mcp-no-toolsearch, mcp-with-toolsearch, mcp-with-code-mode
Results are written to bench/results/. Published results from the study (425 runs) are in bench/published-results/.