CRI

Overview

CRI is a lightweight tool interface used alongside `skill.md` files to run MCP commands more efficiently while dramatically reducing token usage. In the cited examples, CRI acts as a compact execution layer that lets agents access MCP-powered capabilities without carrying the full token overhead of bulky MCP tool schemas in every prompt.

For AI Product Managers, CRI matters because it points to a practical design pattern for scaling agent capabilities without scaling context costs at the same rate. When paired with `skill.md`, CRI can make tool-enabled agents cheaper, faster to orchestrate, and easier to expand across large libraries of skills. This is especially relevant for PMs building coding agents, internal copilots, or workflow automation systems where context efficiency directly impacts latency, cost, and product reliability.

Key Developments

• 2026-01-24: Newsletter coverage highlighted that combining `skill.md` files with a CRI tool can reduce token consumption by more than 70% compared to default MCP setups. The same mention noted that each skill may add only 10–50 tokens to context, making it possible to support far more capabilities than traditional MCP schema-heavy approaches.
• 2026-01-26: AI Jason demonstrated how lightweight skill+CRI integrations can replace bulky MCP agent tools, cutting token usage by over 70% while extending coding agent capabilities at scale.

Relevance to AI PMs

• Lower inference cost and context pressure: CRI offers a concrete way to reduce prompt overhead in tool-using agents, which can improve unit economics and help teams stay within model context limits.
• More scalable capability packaging: By pairing CRI with `skill.md`, PMs can structure large sets of agent skills without exposing full MCP schemas every time, enabling broader feature expansion with less prompt bloat.
• Better architecture decisions for agent products: CRI gives PMs a useful pattern for evaluating whether tool access should be mediated through lightweight abstractions rather than directly through verbose tool definitions.

Related

• ai-jason: AI Jason is the creator/source most directly associated with explaining and demonstrating CRI in the provided mentions.
• mcp: CRI is positioned as a more token-efficient interface for executing MCP commands, rather than replacing MCP capabilities entirely.
• skillmd: `skill.md` works alongside CRI to describe skills in a compact form that minimizes context usage.
• mcp-porter: MCP Porter is described as an open-source bridge that enables agents to run MCP commands via CRI, helping convert existing MCP tools into more efficient skills.