Supabase docs: SSH for AI agents

Supabase has started serving its documentation over SSH. That sounds like a joke from an earlier internet, but it is real, public, and live now at `supabase.sh`. The point is not nostalgia. It is to make the docs easier for AI coding agents to use in the way they already work: from a shell, with simple commands, against a file tree. Supabase says every doc page is exposed as a Markdown file in a virtual filesystem at `/supabase/docs`, so an agent can `ls`, `grep`, `find`, and `cat` its way through the docs instead of relying on a search box. (supabase.com) That matters because coding agents are still bad at knowing what they do not know. Supabase’s own explanation is unusually blunt. Without current documentation, agents guess. They invent APIs, suggest the wrong CLI commands, and skip setup steps like enabling row-level security. Search can help, but search assumes the model already knows what to ask. A shell lets it browse first and decide later. That is the core idea here: give the bot the same interface it uses to inspect code, then point that interface at the docs. (supabase.com) The company is not treating this as a one-off trick. Supabase has been building a broader stack for AI-assisted development. Its MCP server already lets tools like Cursor, Claude, and Windsurf connect to Supabase projects and perform actions through a standard protocol. The docs describe MCP as a way for AI assistants to interact with and query projects on a user’s behalf. Supabase’s feature page goes further. It pitches MCP as a bridge that lets AI tools create projects, design tables, run SQL, manage branches, and pull logs. (supabase.com) Seen in that context, SSH docs are not a strange side project. They fill a gap in the workflow. MCP is for acting on a real project. `supabase.sh` is for getting the agent grounded before it acts. Supabase even provides a command that appends instructions to an `AGENTS.md` or `CLAUDE.md` file so an agent is told to check the docs before making changes. The company also publishes “Agent Skills” and curated prompts for AI tools, which are meant to package task-specific knowledge so models can load it when needed instead of improvising from training data. (supabase.com) The technical detail that makes this less reckless is that Supabase is not exposing a real shell on a public server. The blog and GitHub repo both say the system runs on Vercel’s `just-bash`, an emulated Bash environment written in TypeScript. Commands execute inside a sandboxed JavaScript runtime over a virtual filesystem. In other words, the company kept the familiar shape of SSH without handing anonymous users an actual machine. That design choice is the quiet point of the whole release. Developer platforms want to meet agents where they are, but they do not want to do it by opening a hole in production infrastructure. (supabase.com) Supabase is also making a sharper claim than “AI support” usually means. Most platform AI features still boil down to chat, autocomplete, or a search endpoint wrapped in a nicer UI. This is different. It assumes the agent is not just answering questions but navigating an environment. That is why Supabase argues bash beats search here. Search returns an answer. Bash gives a model room to explore, skim, and follow a trail across files without stuffing its context window with everything at once. (supabase.com) The result is oddly concrete. A developer can point an agent at Supabase, let it authenticate to the MCP server for project actions, and also give it shell-style access to the latest docs for reference. The docs over SSH launched on April 1 as a free public experiment, and the repo that powers it is open source under Apache 2.0. The first command is as plain as the idea itself: `ssh supabase.sh`. (supabase.com)