This is a copy and paste from one of our users.
Hey folks, we use Miro for architectural diagrams and would benefit from an official CLI. Agents do best when they write code [1]. The shell provides a mechanism for composing tools in code via the shell language and pipes. Agents utilize both general purpose languages like Python [6], and DSLs like jq [4], to compose pipelines which can then be saved as new utilities.
CLI tools help to manage context in multiple ways [2][3]:
- By enabling the composition of output into filters [5], the agent is not forced to consume the full output of a tool call, conserving the context window.
- By allowing progressive discovery of documentation via --help and man pages, the agent only fills its context with documentation when it needs to utilize the tool. This is similar to the skills mechanism included in some harnesses.
Since the CLI is utilized by both the agent, humans, build pipelines, etc.; any workflows or utilities utilized by the agent can be incorporated into a deterministic workflow.
A CLI could be automatically generated from OpenAPI specifications, in the style of gws [7], which provides a discoverable command hierarchy and encapsulates the specific REST calls.
[1]: https://machinelearning.apple.com/research/codeact
[2]: https://lucumr.pocoo.org/2025/8/18/code-mcps/
[3]: https://blog.cloudflare.com/code-mode/
[4]: https://jqlang.org/
[5]: https://www.cs.dartmouth.edu/~doug/reader.pdf
[6]: https://arxiv.org/pdf/2503.17181v1
[7]: https://github.com/googleworkspace/cli
