Skip to main content

Manual Quickstart

Four commands take you from a fresh install to an agent that loads exactly the skill it needs during a task. This page runs them against a real catalog — devrelaicom/midnight-expert-tome, thirteen plugins of Midnight development expertise — so you can compare your output with a working setup. If you haven't installed Tome yet, start with Install.

Want to run all of this in one guided step instead? tome init walks you through the same flow and stops to confirm before each action.

1. Add a catalog

A catalog is a git repository of plugins. Add it once, and Tome clones it, parses every plugin inside, and registers them all:

tome catalog add devrelaicom/midnight-expert-tome

Tome confirms the add with the ref it pinned and the plugin count — for this catalog, plugins: 13. Nothing is indexed yet: plugins start disabled, and you choose which ones to enable.

2. Enable a plugin

Enabling a plugin parses, embeds, and indexes its entries for search. Plugins are addressed as <catalog>/<plugin>:

tome plugin enable midnight-expert/midnight-verify

The first enable offers to download the search models if they're missing (see Install). Then list your plugins:

tome plugin list
| Catalog | Plugin | Version | Status | Entries | Last indexed | Last upstream change |
|-----------------|-----------------------|---------|------------|-----------------------------------|--------------|----------------------|
| midnight-expert | midnight-verify | 0.13.0 | ✓ enabled | (19 skills, 2 commands, 7 agents) | just now | just now |

That's one row of thirteen — the other twelve plugins stay ✗ disabled until you enable them.

3. Point a harness at Tome

Tome writes native configuration for each supported harness — rules files, MCP server wiring, and (where the harness supports them) native agents and hooks.

tome harness use cursor

Replace cursor with any of the seventeen supported harness names — claude-code, codex, gemini, zed, copilot, and more. Run tome harness use with no names to configure every harness Tome auto-detects. See Harnesses for the full table and what Tome writes for each.

Run a semantic search across every enabled skill and command:

tome query "verify a Compact contract"
top_k=10 rerank=false min_score=none (10 results)
| Score | Catalog | Plugin | Name | Type | Version | Path |
|---------|-----------------|-----------------|-------------------------------------------|-------|---------|-----------------------------------------------------------|
| 0.7412 | midnight-expert | midnight-verify | midnight-verify:verify-by-execution | skill | 0.13.0 | skills/midnight-verify:verify-by-execution/SKILL.md |
| 0.6685 | midnight-expert | midnight-verify | midnight-verify:verify-by-zkir-checker | skill | 0.13.0 | skills/midnight-verify:verify-by-zkir-checker/SKILL.md |
| 0.6493 | midnight-expert | midnight-verify | midnight-verify:verify-compact | skill | 0.13.0 | skills/midnight-verify:verify-compact/SKILL.md |

(Top three of ten rows shown. The dim header line above the table — shown only in an interactive terminal — reports the effective top_k, rerank, and min_score knobs plus the result count.)

What just happened

That query ran a KNN vector search over the embeddings, entirely on your machine. Reranking is off by default; add --rerank to run the reranker over the KNN hits (it needs the ~563 MB reranker model). Inside a configured harness the same search runs over the MCP server — the agent searches, then loads only the top result instead of holding all 28 indexed entries in context.

Pitfalls

  • Adding a catalog that's already registered exits 4 (catalog_already_exists).
  • Enabling a plugin that's already enabled exits 21 (plugin_already_in_state).
  • A plugin with only a legacy Claude Code plugin.json isn't loaded — that's exit 80 (plugin_not_converted); tome plugin convert migrates it.

Next steps