Supabase Docs over SSH

AI coding agents that operate on outdated information are unreliable, often recommending non-existent APIs or executing incorrect commands. Supabase is addressing this problem not by building a more sophisticated search tool, but by exposing its documentation through a public SSH server, an interface designed for direct exploration.

Supabase has launched supabase.sh, a service that provides AI coding agents with direct bash access to its documentation. (Source). The core issue is that agents without access to current information tend to guess, leading to errors. The new service presents the same markdown source files that power the supabase.com/docs website, but through a command-line interface. The system is built on just-bash, an open-source Bash emulator from Vercel that is written in TypeScript. This architecture allows commands to run inside a sandboxed JavaScript runtime over a virtual file system. This design choice means the service can remain public and unauthenticated without exposing an actual shell on the server.

The strategic decision here is the choice of interface. Instead of a more conventional and opaque search API, Supabase has opted for bash. This is significant because it shifts the interaction model from querying to exploration. A search tool requires a predefined query and returns a set of results, limiting an agent's ability to browse or discover context. Bash, by contrast, allows an agent to list pages, search across multiple files, and read portions of a document to assess relevance before committing to the full content. This mirrors the workflow an individual might use when navigating a local codebase, an environment AI coding agents are extensively trained on.

This approach treats documentation as a navigable file system rather than a database to be queried. It meets the agents in their native environment. The implementation, leveraging Vercel’s just-bash, also points to a pragmatic strategy of assembling existing open-source components to create a novel service. The innovation is not in the content, which remains the markdown source, but in the directness and familiarity of the access method. By making the project open source, with its code available at github.com/supabase-community/supabase-ssh, Supabase is transparent about this construction.

The open question this experiment poses is about the future of agent-tool interaction. If providing direct, explorable access to a source of truth via a familiar, low-level interface proves more effective for AI coding agents than a high-level, abstract API, the implications are considerable. It suggests that for certain use cases, the most effective way to ground an agent is not to build a better black box, but to give it a key to the library.