Skip to content

Agent skills ​

AutoRAG ships Markdown skills that a coding agent can read to install, configure, search, or diagnose AutoRAG. This page lists what each skill does, how to copy it into an agent, and the sequence an agent follows to answer a question.

Shipped skills ​

Each skill is a SKILL.md with YAML frontmatter (name, description, license) and a body of rules and commands.

SkillDirectoryPurpose
autoragskills/autorag/Search, summarize, compare, and answer from an already configured model-backed librarian.
autorag-setupskills/autorag-setup/Install AutoRAG and configure the model-backed librarian: the single search model, approved roots, indexes, datasources, health checks.
autorag-lite-setupskills/autorag-lite-setup/Install and register the model-free autorag-mcp server, write model-free config, build indexes, and verify MCP search.
autorag-doctorskills/autorag-doctor/Diagnose and repair a broken install: probe every datasource natively, repair locks, embedding mismatches, and stale indexes, and finish with a per-datasource status table.

Pick the skill by intent. Use autorag-lite-setup for external-agent search over MCP with no search model; use autorag-setup (plus autorag) only when the agent must also drive the model-backed librarian. Use autorag-doctor only to repair an existing install. See Choosing an integration for the full decision table.

Installing skills ​

Install the skills into your coding agent ​

Skills are not auto-discovered. From a clone of the repository, copy the skill folders into the agent's skill directory. Claude Code reads .claude/skills; other agents use their own directory (for example Cursor, Codex, and Pi each have a per-agent or project skill location), so copy the same folder there and reload the agent session.

bash
mkdir -p .claude/skills
cp -R skills/autorag-lite-setup .claude/skills/

Copy skills/autorag, skills/autorag-setup, and skills/autorag-doctor too when the agent should drive the model-backed librarian or repair an install:

bash
mkdir -p .claude/skills
cp -R skills/autorag-lite-setup skills/autorag-setup skills/autorag-doctor .claude/skills/

The npm package ships the same skills/ folder. A global install keeps it under the package root:

bash
AUTORAG_SKILLS="$(npm root -g)/@autorag/librarian/skills"
mkdir -p .claude/skills
cp -R "$AUTORAG_SKILLS/autorag-lite-setup" .claude/skills/

With a Bun global install the package root is ~/.bun/install/global/node_modules/@autorag/librarian/skills.

After copying, reload the agent so it picks up the skill, and run the setup skill once to register the stdio MCP server with the host. Do not copy the removed autorag-lite-search skill: routine Lite retrieval goes through MCP tools, not a shell command.

Skills in the npm package ​

package.json files publishes dist, skills, internal-skills, README.md, NOTICE, licenses, and vendor.

  • skills/ holds the parent-agent skills above (autorag, autorag-setup, autorag-lite-setup, autorag-doctor).
  • internal-skills/ holds the librarian agent's own turn rules, for example internal-skills/organize/. When Jev routes a request to organize, the run loads that file as a skill invocation. These are never parent-agent skills: do not copy internal-skills/ into another agent.

Plugin marketplaces ​

Three marketplace manifests declare one plugin, autorag-remote, sourced from plugins/remote-agents/: .claude-plugin/marketplace.json, .cursor-plugin/marketplace.json, and .agents/plugins/marketplace.json. They ship the cloud-agent search skill, not an MCP server; you register your own server URL. Install from the marketplace:

bash
claude plugin marketplace add Marker-Inc-Korea/AutoRAG
claude plugin install autorag-remote@autorag
bash
codex plugin marketplace add Marker-Inc-Korea/AutoRAG
codex plugin add autorag-remote@autorag

Then follow Always-on cloud agents to start the HTTP server and add its URL.

Using the skills ​

Quick agent workflow ​

  1. Confirm the CLI and config exist. command -v autorag should print a path; a config is resolved from --config, AUTORAG_CONFIG, $AUTORAG_HOME/config.json, or ~/.autorag/config.json. If the CLI is missing, install it: command -v autorag >/dev/null || bun install -g @autorag/librarian.
  2. Check readiness. autorag status --json reports corpus freshness and index health; autorag duplicates --json scans duplicate families read-only.
  3. Build or refresh indexes if status is stale. autorag refresh --json runs all configured methods incrementally; autorag refresh --method minsync --json narrows the run. Indexing never happens while answering.
  4. Search. For the model-backed librarian use autorag search "your question" --json (add --debug for sessionId and diagnostics); for Lite use the MCP tools autorag.status → autorag.refresh (when needed) → autorag.search.
  5. Persist and re-open evidence when the caller needs exact chunks. See Evidence and citations.
  6. On failure, read the diagnostics. autorag health --json checks the model and credentials; the autorag-doctor skill maps diagnostic codes to first moves.

Safety rules every skill states: setup, indexing, and refresh never move, rename, edit, or delete source documents; never permanently delete (OS trash only); never print or leak credential values (record environment-variable names instead); and index artifacts go only under the workspace .autorag/ directory and Jikji's central store.

Released under the Apache-2.0 license.