Install

One token is all you need. Step ① alone is enough for an agent to read and write.

Get a token first

Issue one at /settings/tokens and the snippets in ① below come back with the token already filled in, one per client. The token is shown that once and never again — the server keeps only a sha256 hash, so if you lose it you issue a new one.

On Claude Code it is two lines

The plugin brings the server config and the skills together. These two lines stand in for ① and ② of the three steps below, so only ③ is left.

Run these in order inside Claude Code
/plugin marketplace add neruu00/cushion
/plugin install cushion@cushion

It asks for the token during install. Paste it and it goes into Claude Code's secure storage, so it never sits in a config file in plain text — which is what makes this safer than the command in ①.

Skills that arrive through a plugin are prefixed with the plugin name. In ③, use /cushion:cushion-use instead of /cushion-use. The rest follow the same shape, like /cushion:cushion-new.

If you use Antigravity or Codex CLI, or would rather not install a plugin, the three steps below do the same job. Both routes connect to the same server.

Connecting in three steps

The Cushion server is a Streamable HTTP server that needs nothing but a URL and an Authorization: Bearer header, so it works with any client. Pick yours in ①. Step ③ is the same whichever client you use.

  1. 1

    Connect

    Run it once in a shell. The doc_* tools show up from then on.
    Run once in a shell
    claude mcp add --transport http --scope user cushion https://www.cushion.pics/api/mcp --header "Authorization: Bearer cshn_pat_…"
  2. 2

    Install the skills

    Optional, but recommended. The basics work without them.

    Every project on this machine

    Putting them in ~/.claude/skills/ installs them once for every project. Skill files hold nothing project-specific, so this is usually the right choice.

    Send this to your agent as is (~/.claude/skills/)
    Fetch https://www.cushion.pics/api/skills and save each entry in files to ~/.claude/skills/<path>. Overwriting existing files is fine.

    This project only

    Put them in .claude/skills/ and commit, and your whole team gets them automatically. You install per project, and a committed copy stays put even when the original moves on.

    Send this to your agent as is (.claude/skills/)
    Fetch https://www.cushion.pics/api/skills and save each entry in files to .claude/skills/<path>. Overwriting existing files is fine.
  3. 3

    Pick which library to use

    One line in AGENTS.md, and the next session starts from it.
    Send this to your agent as is
    /cushion-use

    The agent looks up git remote to find the library for this project and writes it into AGENTS.md. Skip it if a teammate already committed that line.

③ is the least visible and the most important

Having no docs in the repo is the normal state here, so without a marker an agent simply walks past. That one line in AGENTS.md is the only signal that makes an agent find Cushion, and /cushion-use writes it for you.

If a teammate already committed AGENTS.md, anyone joining only has to do ①.

Rolling it out to a team

For yourself the three steps above are the whole story. To commit it to the repo and hand it to everyone, grab the .mcp.json snippet from the library settings. That file never holds the token. It gets committed, so it uses environment variable expansion (${CUSHION_TOKEN}).

Project-scope config overrides user scope. If you were already connected and then pick up only .mcp.json, you get a 401 inside that project — and only there — until CUSHION_TOKEN is set.