aclif

aclif — "agent CLI framework" — builds command-line tools meant to be run by AI agents rather than by people. It is MIT-licensed, published on npm as @aclif/core, and developed by Prompt One, Inc.

The problem it targets

An MCP server publishes a fixed list of tools, and each tool's definition occupies the agent's context on every turn. That forces the server's author into a trade at build time: publish every operation and pay tokens for all of them continuously, or publish a handful of broad ones and put the rest of the API out of reach. An agent spanning five platforms also inherits five processes, five logins, five grammars and five error formats.

aclif inverts the default. A command's schema, examples and safety metadata load only when the agent asks for them, so the whole API of every bundled provider stays reachable at no standing context cost, and one grammar and one error vocabulary cover every provider.

Introspection before execution

Every command can be inspected without running it. --schema returns flags, arguments and safety metadata; --examples returns runnable examples with the responses they produce; --dry-run prints what the command would do. Each command declares ACI metadata up front — mutability, idempotency, reversibility, blast radius, API calls consumed, whether confirmation is required — so a caller can judge a command before committing to it.

Errors are built for recovery rather than for a human reader: a failed call returns a structured envelope naming every way to supply what was missing, with a distinct exit code.

Three deployment modes

The same binary runs spawned by an agent that holds its own credentials, embedded in a host application that supplies them, or behind a gateway serving many agents with centralised policy and audit. Because a command is just a string, a workflow author can fix the exact command at design time and have the agent execute it with no tool definition loaded and no inference at all.

Providers

Native support covers Salesforce, ServiceNow, DocuSign and Agentforce, with Google Workspace (Gmail and Calendar) contributed. Custom providers can be built against the documented contract.