The agent interface spec that undercuts MCP

What it does

AXI is a specification and benchmark suite defining ten principles for “agent-native” CLI tools—interfaces designed for LLMs rather than humans. The repo includes reference implementations like gh-axi for GitHub and chrome-devtools-axi for browser automation, plus harnesses that compare them against raw CLIs and MCP servers.

The interesting bit

The benchmarks do the talking. Across hundreds of runs with Claude Sonnet 4.6, the AXI wrappers consistently undercut MCP on token cost while matching or beating its success rate. The trick is aggressively minimal output: TOON instead of JSON, truncated text with --full escape hatches, and pre-computed aggregates that eliminate round trips.

Key highlights

• Browser benchmark claims 100% success at $0.074 per run versus $0.101–$0.120 for MCP variants; GitHub benchmark claims 100% success at $0.050 versus $0.148 for GitHub MCP
• Ten principles include “content first” (running with no arguments shows live data, not help text) and “contextual disclosure” (suggest next steps after every output)
• Reference implementations are thin, opinionated wrappers around existing tools like the official gh CLI and chrome-devtools-mcp
• Growing catalog of community AXIs covering Slack, SQLite, Google Workspace, and Metabase
• Benchmark harnesses and raw study data live in the repo, not just summary tables

Caveats

• The reference implementations are wrappers, not standalone protocols; they inherit fragility from the underlying tools they reshape
• Published benchmarks were run exclusively with Claude Sonnet 4.6, so mileage may vary with other models

Verdict

Study this if you’re building agent tools and suspect MCP is burning tokens on ceremony. Skip it if you need a replacement wire protocol—AXI is currently a design philosophy and wrapper layer, not a low-level standard.