Glossary

An open standard for how a command-line tool behaves when an AI agent runs it. The name also refers to the mode itself: the stable, schema-governed output a tool emits when it's invoked with --output jsonl.

Describes a tool an AI agent can run, read and act on without guessing, because its output is typed, its capabilities are discoverable, its reads are bounded, its defaults are safe and it states its own outcome.

Agent-Operable Interface: the formal specification behind Machine Mode. Each interface surface gets its own profile.

The command-line profile of AOI, currently v1.0-alpha. It defines machine-mode output, discovery, errors, completion and composition for CLIs.

A tool's default, human-facing output: tables, color, progress bars and prose. AOI-CLI leaves it untouched and adds machine mode beside it.

JSON Lines: one complete JSON object per line, UTF-8, every line ending in a newline. It's the wire format of AOI-CLI machine mode.

One JSON object on one line of machine-mode output. Every event has a type field; framework events use the aoi: prefix and domain events belong to the tool.

The reserved aoi object on every event. It holds framework fields such as seq and run, so they can never collide with a tool's own field names.

The first event of a run. It names the tool and carries the tool version, the AOI version and the schema name and version.

The last event of a finite run. Its ok field states whether the tool did what was asked, with no default; a stream that ends without it is incomplete.

A structured error event with a category from a closed list, a stable code, a message and a required retryable flag.

A required boolean on every aoi:error. True means a retry may succeed, optionally after retry_after_ms; false means it won't, or that the tool can't tell whether a side effect already happened.

A per-run counter in the envelope that starts at 0 and has no gaps. Checking it against aoi:summary.event_count shows whether lines went missing from the middle of a stream.

The schema and capabilities commands, which describe a tool with no credentials, no network access and no side effects.

The output of the capabilities command: every command a tool offers, flagged read_only, destructive, requires_confirm or supports_idempotency_key as applicable.

A tool's position in a pipeline. A source emits JSONL, a transformer reads and emits it, and a sink reads it and performs side effects. Each still emits aoi:meta first and aoi:summary last.

One of the ten properties of an agent-operable tool: Typed, Discoverable, Streamable, Bounded, Safe, Idempotent, Auditable, Verifiable, Composable and Versioned.

A pattern in which an AI agent writes a short program that calls tools, instead of making one tool call per model turn. Cloudflare named it in September 2025.

Model Context Protocol: the protocol agents use to discover and call tools exposed by servers. A Machine Mode tool can sit behind an MCP server.

A measurement tool that runs any CLI and reports what its machine output actually does, without assuming the tool follows AOI. It produced the site's survey.

The conformance checker the spec defines. It reports conformance per factor rather than as a single level, and is in development.