Claude Code

Give Claude Code the skill, the MCP server, the stop hook, the language server, and commands in one install, with the Ascribe plugin.

Claude Code takes Ascribe in one of two ways: the plugin, which gives it everything in one install, or the project’s own entries, written by ascribe agents sync. Use one, not both: with both, it hears about each problem twice.

The plugin

The Ascribe plugin, in plugins/ascribe/ in Ascribe’s repository, gives Claude Code:

• The skill: how to write, check, and fix pages.
• The MCP server, whose tools are for an agent with no shell; one with a shell does as well with the commands.
• The stop hook, and the language server, which Claude Code starts for .md files and whose diagnostics it reads after each edit: they check for an agent that doesn’t check its own work. The language server reports the editor’s build, as the edit hook does, and only for edits Claude Code makes with its own tools, not through its shell.
• Commands: /ascribe:check, /ascribe:new-page, and /ascribe:review, which run the named prompts. /ascribe:new-page takes what the page is for: the agent chooses the page type from ascribe model, then runs the prompt.

Add it from a copy of the repository, in Claude Code in a terminal (its VS Code extension doesn’t offer /plugin):

/plugin marketplace add ./plugins/ascribe
/plugin install ascribe@ascribe

The plugin runs the ascribe binary the VS Code extension would (see Which ascribe it runs), with an environment variable in place of the setting: the one ASCRIBE_PATH names, then the project’s node_modules/.bin/ascribe, then ascribe on your path. So a project that pins its version with npm install --save-dev @ascribed/cli gets that version in every tool, whatever is installed globally.

The plugin has both kinds of manifest: .claude-plugin/plugin.json, which Copilot’s CLI reads too, and the Agent Plugins plugin.json, which VS Code and Cursor read. To keep the language server’s diagnostics out of the conversation on a project with many problems, see the plugin’s README; the stop hook still checks before Claude finishes.

The project’s entries

Claude Code doesn’t start a plugin’s language server in a cloud session. There, use the project’s entries, which Claude Code reads from the repository:

ascribe agents sync --target claude --with-hook

That writes the skill where Claude Code reads it, .claude/skills/ascribe/, an import of AGENTS.md in CLAUDE.md, and the hooks and the MCP server in .claude/settings.json and .mcp.json.

Claude Code reads AGENTS.md by itself only while there’s no CLAUDE.md at or above the folder it works in. So sync never creates a CLAUDE.md unless you ask with --target claude, and when there is one, it adds the import to it rather than copying the rules. A CLAUDE.md that imports AGENTS.md itself, with its own @AGENTS.md line, is left as it is, so it doesn’t load the file twice. A personal CLAUDE.local.md also stops Claude Code reading AGENTS.md; sync doesn’t edit it, but says so, and you can add @AGENTS.md to it yourself.