wevm/incur

What it solves

incur is a framework for building command-line interfaces (CLIs) specifically optimized for AI agents. It solves the problem of high token consumption and discovery friction when agents interact with software tools, reducing token usage by up to 3x per session compared to traditional MCP or skill-based alternatives.

How it works

It provides a lightweight API to define commands, arguments, and options using schemas (via Zod). It optimizes the agent experience through:

• Agent Discovery: Built-in commands (skills add, mcp add) allow agents to automatically find and register the CLI.
• Token-Efficient Output: Defaults to the TOON format, which strips redundant syntax (like braces and quotes) to reduce token costs by up to 60% compared to JSON.
• Call-to-Actions (CTAs): Allows developers to suggest the next relevant commands to an agent after a run, enabling seamless operation chaining.
• Flexible Deployment: CLIs can be served as standalone binaries, standard Fetch API handlers (for Bun, Deno, Cloudflare Workers), or as MCP servers over HTTP.

Who it’s for

Developers building tools, APIs, or internal utilities that they want to make easily discoverable and highly efficient for LLM-based agents to use.

Highlights

• Token Savings: Significantly reduces tokens spent on session start, discovery, and responses.
• ** uma MCP Integration:** Supports both stdio and HTTP-based Model Context Protocol (MCP) servers.
• Multi-Modal Interface: The same codebase can function as a CLI, a Web API, and an MCP server.
• Schema-Driven: Uses Zod for well-formed I/O and automatic type inference.
• File-Based Routing: Supports loading commands from sibling files and folders for better project organization.