Giving AI agents design system context from the terminal: what we learned building a CLI
Atlassian expanded AI agent access to its design system by launching a command-line tool, complementing existing MCP server and skill methods to improve reach and performance.
Atlassian introduced a command-line interface (CLI) to distribute its design system guidance to AI agents, adding to existing MCP server and skill-based methods. The team chose to build the CLI because it broadens access to agents without requiring an MCP client, while maintaining data consistency and reducing token usage and latency. The CLI leverages the same structured content—component APIs, design tokens, icons, accessibility guidance, and lint rules—used by the MCP server and skill, ensuring no divergence in guidance across entry points.
The CLI was developed in under a week by reusing existing infrastructure, including typed schemas for content and shared modules for search and ranking logic. This approach eliminated redundant documentation maintenance and allowed the CLI to focus on terminal-specific needs such as argument parsing, human-readable output, and JSON support for scripts and agents. By consuming the same shared modules as the MCP server and skill, the CLI avoided duplicating data or handlers, preventing drift between distribution methods.
After the initial release, the team spent a month benchmarking and refining the CLI based on real-world usage. Key improvements included reducing startup overhead by integrating the CLI into Atlassian’s internal atlas CLI, adding a batch command to consolidate queries, and enhancing fuzzy search capabilities. These changes targeted inefficiencies observed in early agent traces, where multiple round trips and slow lookups impacted performance.
To validate the CLI’s effectiveness, Atlassian ran internal benchmarks comparing agent performance across different context-fetching methods: MCP server, CLI, CLI wrappers, and bundled reference files. Metrics included task pass rates, time per task, token usage, and tool calls, with transcripts revealing whether agents followed intended instructions. The benchmark confirmed the CLI’s viability, particularly after reducing overhead from over 15 seconds per query with npx to milliseconds via the internal CLI.