Skip to content

How MCP delivery works

Quiver delivers skills only through its MCP server, quiver mcp. The assistant starts it on stdio at the beginning of a session.

The server serves the project of its working folder, found by walking up to .quiver.json, or of $QUIVER_PROJECT when that is set. Without a project it serves the global skills only. It logs the folder and the project to ~/.quiver/logs/mcp.log.

MCP servers can send instructions when a session starts. Quiver's instructions are short. They tell the assistant:

  • how to use the tools: get_skill, get_skill_file, list_skills and search_skills,
  • that new skills are saved with create_skill,
  • the names of the always-on skills, for an assistant without a session-start hook,
  • to call list_skills before its first reply when the conversation has no <quiver-skills> block.

The skills themselves come from a session-start hook, which quiver install sets up for Claude Code, Codex, Cursor, Antigravity and VS Code. It adds the always-on skills in full, and the ID and description of every other skill served here, project skills first, to the conversation. An assistant takes only so much text from one hook command, so Quiver splits it over as many hook commands as needed. A skill that cannot be read is named in a <quiver-notice> block. See Always-on skills.

Tool descriptions are not used for the catalog: Claude Code defers MCP tools, so the model does not see their descriptions until it looks them up.

When a task fits a skill's description, the assistant calls get_skill with its ID. The result is the skill's instructions, a list of its extra files, and the skill's folder. The assistant reads an extra file with get_skill_file, and runs a skill's scripts itself, from that folder.

Only the skills a task needs are read into the conversation.

Every served skill is also an MCP prompt. In Claude Code it appears as /quiver:<id>. Text after the command is passed to the skill as the task.

While a session runs, the server watches the library and the assignments. Skills you add, remove or change show up in list_skills and the prompts right away; the server sends prompts/list_changed. The hook's text is built when a session starts, so the skill list and the always-on skills in it change from the next session.

The assistant can also write to your library:

  • create_skill creates a local skill, commits it and assigns it to this project or globally.
  • update_skill changes a local skill served in this project and commits the change. Source skills are read-only, and so are always-on skills: change those with quiver edit.
  • search_skills searches the whole library, also skills not assigned here. The assistant may suggest an unassigned skill, but asks before using it.

See MCP tools and prompts for the inputs and results of every tool.

Not every assistant passes instructions to the model up front. Claude Desktop drops them and has no hook, so there skills are used when you ask for them. The other supported assistants run the session-start hook, so the skills reach the model without the instructions; see Always-on skills and Connect an assistant.