Spaces: A CLI Built for Humans and Agents

Mistral AI published an engineering blog post describing Spaces, an internal CLI its solutions team built to ship faster. Spaces scaffolds projects, spins up development environments, generates configs, and deploys to staging. The post says a typical session consists of three commands — spaces init my-project, cd my-project, and spaces dev — taking a project from nothing to a running multi-service setup with hot reload, a database, and generated Dockerfiles.

The post states that as the CLI's scope and internal user base grew, the team found it was building for coding agents as well as human developers. An agent that tried to use the tool's TUI module picker saw raw ANSI escape codes and could not send arrow keys or toggle selections, locking it out of the command. The team addressed this by giving every interactive prompt a flag equivalent, with smart defaults for headless mode under a -y flag.

The post describes a plugin system in which each component declares its own properties, such as type_id, category, default_port, environment variables, and dev command. Plugins are introspectable: a human browses a TUI picker while an agent queries the registry and receives JSON. Adding a module type now means writing one plugin class instead of updating the picker, Dockerfile generator, env file writer, and compose template.

On every init, Spaces generates two files: context.json, a structured snapshot of modules, ports, commands, and env vars, and AGENTS.md, rules written for LLMs. The context file updates automatically on the next dev or init. The post also describes replacing a current-working-directory dependency in the add command with an explicit path parameter that falls back to searching parent directories.

In a practical example, the post says an agent was retroactively asked to wire a repository into the Spaces CLI for deployment. It ran --help across relevant commands, generated a config.yaml, connected Dockerfile and registry settings, and set up a GitHub Actions CI pipeline. From a single prompt to a live deployment took under 10 minutes, a cycle time the team says it is working to reduce. The interactive demos embedded in the blog post are deployed with Koyeb as a Space scaffolded and deployed through Spaces itself.

The post credits Lorenzo Signoretti, Riwa Hoteit, and Sam Fenwick at Mistral AI, thanks the Applied AI team, and says the company is hiring for developer tools at the intersection of AI and infrastructure.