aclif

aclif (Agent CLI Framework) is an MIT-licensed, open-source TypeScript project by Prompt One, Inc. that builds command-line tools purpose-built for AI agents. Published on npm as @aclif/core and built on top of oclif (the framework behind the Salesforce and Heroku CLIs), it gives agents a single binary that covers every provider it ships with, using one grammar, one JSON envelope, and one error vocabulary across all of them.

What It Is

aclif is a CLI framework designed to solve a specific problem in agentic workflows: the cost and complexity of connecting AI agents to multiple SaaS platforms. Rather than requiring a separate MCP server, login, grammar, and error format per platform, aclif provides a single command surface where a command's schema, examples, and safety metadata load only when the agent asks for them — so nothing sits in context by default. The framework is installable via npm install -g @aclif/core and requires Node.js 22 or later.

Why It Differs from MCP

The project's documentation draws a direct contrast with the Model Context Protocol (MCP) server approach. With an MCP server, every tool on the published list occupies the agent's context on every turn, forcing the server author to trade coverage for token cost at build time. aclif inverts this: one tool describes the grammar for all providers, and a command's definition loads only when the agent requests it with --schema or --examples. This means the full API surface of every provider is reachable at no standing cost in context. Additionally, for defined workflows, the exact command can be determined once at design time and embedded as a string — at runtime the agent executes it as ordinary code with no model in the loop and zero inference cost.

Core Capabilities

• One grammar: A single command structure, JSON envelope, and error vocabulary across every provider. Adding a new platform adds commands without adding grammar.
• Canonical names: Alias sets map logical names like customer to Account in Salesforce or core_company in ServiceNow; --canonical resolves them at runtime.
• Actionable errors: Every error names the failure, the command that fixes it, and where possible a correctedValue and workingExample ready to resend — enabling single-turn recovery.
• Introspection without execution: Seven flags (--schema, --examples, --shape, --changelog, --discover, --flags-for, --estimate) return before the command runs, require no credentials, and consume no API quota.
• Declared safety metadata: Mutability, blast radius, reversibility, and idempotency are declared on every command. Every mutation accepts --dry-run and demands --confirm where metadata requires it.
• Embeddable runtime: The same command classes run in-process inside a host application or gateway, keeping connection pools warm and caching logins per instance.

Three Deployment Models

aclif supports three distinct runtime topologies:

• Run by the agent — the agent spawns the binary directly; credentials come from flags, environment variables, or a config profile. Best when one agent, one operator, and one credential set share a trust boundary.
• Run by a host application — an application sits between the model and aclif, holds credentials, and calls runtime.run(...) in-process. Best when the model must never hold credentials and tool definitions must stay out of its context.
• Run by a gateway — a single long-lived process serves many agents, resolving credentials from an enterprise vault per request, enforcing policy via a capabilityGate hook, and recording every call. Best when many agents share providers and one place must hold policy and audit.

Providers and Extensibility

Five providers ship with the current release: Salesforce (native), ServiceNow (native), DocuSign (native), Agentforce (native), and Google Workspace/Gmail/Calendar (contributed). Native providers are maintained by the project with live smoke tests against real instances; contributed providers are maintained by their authors. A private tier allows forks to keep proprietary providers without upstream conflicts. New providers can be added via a JSON manifest for a single endpoint or a full TypeScript provider generated from an OpenAPI document — the project includes a sample generation prompt and an add-provider skill for Claude Code.

Update: v1.3.1

The latest release is v1.3.1, published on September 16, 2026. The repository was created on September 12, 2026, and CI runs on Linux and Windows under Node 22 and 24. The test suite covers six layers: unit, conformance, provider fixture, end-to-end binary, contract goldens, and an opt-in live smoke test. The project is actively maintained by Prompt One, Inc. under the MIT license.