AXI is ten rules for CLIs that agents run

AXI is a set of ten design principles for CLIs built for agents. The catalog read on 2026-10-06 lists four official reference tools: gh-axi for GitHub, chrome-devtools-axi for browser automation, lavish-axi for human review, and quota-axi for quota and usage. The principles focus on compact output, predictable failures, and useful next steps.

AXI's ten principles

The AXI specification says these ten rules define what makes a CLI an AXI. The wording below follows the summaries in principles.yaml.

1. Token-efficient output. Use TOON format for about 40% token savings over JSON.
2. Minimal default schemas. Return 3 to 4 fields per list item, not 10 or more.
3. Content truncation. Truncate large text with size hints and provide a --full escape hatch.
4. Pre-computed aggregates. Include counts and statuses that eliminate extra round trips.
5. Definitive empty states. Print explicit 0 results instead of ambiguous empty output.
6. Structured errors and exit codes. Use idempotent mutations, structured errors, no interactive prompts, and loud failure on unknown flags.
7. Ambient context. Install opt-in session integrations first, then offer an on-demand skill.
8. Content first. Running a CLI with no arguments should show live data, not help text.
9. Contextual disclosure. Include next-step suggestions after each output.
10. Consistent way to get help. Provide a concise per-subcommand reference when an agent needs it.

These rules target the points where an agent pays for a poor interface. Long JSON consumes context. Blank output can look like a failed command. A mutation that prompts for input cannot finish unattended. AXI makes each state explicit enough for the next shell command to be chosen from the response.

AXI's four official tools

The official catalog has four entries, each showing a different use of the same interface rules.

The principles are visible in the command shapes. chrome-devtools-axi open combines navigation with a snapshot, while its interaction commands can return the updated page state and next-step guidance. gh-axi exposes a live dashboard when run without a subcommand and keeps command guidance in its output. quota-axi emits compact TOON rows for quota, exhaustion, and attention instead of making the caller assemble several provider responses. lavish-axi carries annotations and selected text through a local review session. The four tools apply the rules to different kinds of state.

The first two tools publish benchmark tables. The gh-axi README describes 17 GitHub tasks across five interface setups, repeated five times with claude-sonnet-4-6. It reports 100% task success, 46,462 average input tokens, $0.050 per task, 15.7 seconds, and three turns. The README does not state when those runs happened.

The chrome-devtools-axi README describes 14 browser tasks across seven setups, repeated five times for 490 runs with the same model. It reports 100% task success, 79,141 average input tokens, $0.074 per task, 21.5 seconds, and 4.5 turns. Its README also gives no benchmark run date. These are useful published comparisons, but the missing dates make them point-in-time evidence rather than a dated trend.

Both tables measure defined task suites, not every GitHub or browser job. The READMEs say an LLM judge scores task success. The 100% figures therefore describe the published tasks under their listed conditions. They do not establish a universal success rate for other models, private repositories, authenticated sites, or write operations. The useful comparison is narrower: within the same task set, AXI reports fewer input tokens, lower cost, and fewer turns than the listed alternatives.

quota-axi and lavish-axi

The quota-axi README lists 19 providers: Claude, Codex, Cursor, GitHub Copilot, Grok, Kimi, Z.AI, Alibaba, OpenCode Go, Antigravity, Command Code, MiniMax, MiMo, DeepSeek, OpenRouter, ElevenLabs, Devin, Muse, and Higgsfield. The catalog row lists only five: Claude, Codex, Cursor, Copilot, and Grok. That is a catalog mismatch, not evidence that the executable supports only five providers.

The data-only description has a boundary worth stating precisely. quota-axi says it does not route requests, recommend a provider or model, mint or rotate credentials, or proxy traffic. Its README also documents a narrow case where an expired credential may be handed to the vendor's own non-interactive CLI for renewal. Data-only therefore describes the tool's role, not a promise that it never invokes a local provider command.

On 2026-10-05, version 0.1.58 added read-only Higgsfield credits and jobs support under issue #302. That release is a concrete example of the tool extending its reported data without becoming a router.

lavish-axi is a different kind of reference tool. Its README describes local HTML review, element and text annotations, Mermaid whiteboard editing, and feedback sent back to the agent. The catalog labels it "Human review". It is the review loop in the four-tool set, not another wrapper around an external service API.

tasks-axi and the catalog boundary

tasks-axi is a separate backlog manager from the kunchenguid account, designed with AXI. It edits a hand-written backlog.md in place and preserves a byte-exact round trip, so an agent can change a task without regenerating the whole file. The AXI catalog read on 2026-10-06 lists many projects but does not list tasks-axi in either its official or community table. The tool follows the same design language while remaining outside the published catalog.

Sources

• AXI principles specification: https://raw.githubusercontent.com/kunchenguid/axi/main/principles.yaml (read 2026-10-06)
• AXI catalog: https://axi.md/ (read 2026-10-06)
• AXI core repository: https://github.com/kunchenguid/axi (read 2026-10-06)
• gh-axi repository and benchmark: https://github.com/kunchenguid/gh-axi (read 2026-10-06)
• chrome-devtools-axi repository and benchmark: https://github.com/kunchenguid/chrome-devtools-axi (read 2026-10-06)
• quota-axi repository and provider documentation: https://github.com/kunchenguid/quota-axi (read 2026-10-06)
• quota-axi 0.1.58 changelog: https://raw.githubusercontent.com/kunchenguid/quota-axi/main/CHANGELOG.md (read 2026-10-06)
• lavish-axi review tool repository: https://github.com/kunchenguid/lavish-axi (read 2026-10-06)
• tasks-axi backlog tool repository: https://github.com/kunchenguid/tasks-axi (read 2026-10-06)

Last verified: 2026-10-06.