Agent skills: the LakehouseBox plugin for agents
LakehouseBox publishes one plugin for agents in the public repository lakehousebox/agent-skills (MIT). It holds two things:
- the connector: the read-only MCP server at
https://mcp.lakehousebox.com/mcp. It needs a LakehouseBox account (free), approved once in the browser (an agent with a shell can start without one:lhbox new, /new); - the skills: instructions the agent loads when a task matches (Agent Skills), with references it opens only when it needs them.
| skill | what it teaches the agent |
|---|---|
lakehousebox |
the base: which way in (the lhbox CLI, the connector, or plain DuckDB), reading anyone's public catalog with no account, the rules that save a retry, and where the docs cover what no skill does yet |
lakehousebox-load-data |
files into tables in one commit (CSV, Parquet, GeoParquet), appending the next batch, the file bucket, upload-only keys for devices (import) |
lakehousebox-keep-fresh |
a table that fills itself: scheduled SQL that reads a URL, or an ingest sink a device or webhook posts rows to (scheduled SQL, ingest) |
lakehousebox-share |
access for colleagues and agents per catalog, publishing a catalog and its risks, AGENTS.md and table descriptions (access) |
lakehousebox-hosted-agents |
a prompt LakehouseBox runs on a schedule next to the data, tested before it sends anything (a preview; hosted agents) |
lakehousebox-portolan |
reading a Portolan geospatial catalog hosted on LakehouseBox with no account (from a collection to its table, spatial filters that stay fast, units, the errors and what they mean), and hosting one (guide) |
The base skill is what every LakehouseBox task needs; the workflow skills (loading data, keeping tables filled, sharing, hosted agents) and the use-case skills such as lakehousebox-portolan build on it. Reading a public catalog needs neither an account nor the connector: the skills show the agent how to read it with DuckDB.
Install
| agent | install | checked |
|---|---|---|
| Claude Code | /plugin marketplace add lakehousebox/agent-skills, then /plugin install lakehousebox@lakehousebox |
2026-10-11, Claude Code 2.1.296: two skills and the connector |
| Codex | codex plugin marketplace add lakehousebox/agent-skills, then codex plugin add lakehousebox@lakehousebox |
2026-10-11, Codex CLI 0.162.1: the plugin and the connector |
| Gemini CLI | gemini extensions install https://github.com/lakehousebox/agent-skills |
2026-10-11, Gemini CLI 0.63.0: two skills, the connector and the context file |
| Claude (claude.ai, Desktop) and Cowork | upload lakehousebox.zip and lakehousebox-portolan.zip as custom skills, and add https://mcp.lakehousebox.com/mcp as a custom connector |
not checked yet |
| any agent that reads skills from a folder | copy skills/<name>/ from the repository into that agent's skills directory |
not checked yet |
After installing, the person approves the connector once in the browser; an agent that only reads public catalogs never needs it.
What it changes
A fresh Claude Code agent, with no account and no connector, was asked which barrio of Madrid holds Puerta del Sol and how many tourist flats it has, from the public catalog portolan/madrid (2026-10-10). Both answers were right (Sol: 8 records, 14 flats):
| with the skills | without | |
|---|---|---|
| turns | 7 | 15 |
| time | 34 s | 109 s |
| cost reported by the agent | $0.27 | $0.61 |
Without the skills the agent found an old DuckDB on the machine, built a newer one, and wrote its spatial join without the filter that lets DuckDB skip data.
How the skills stay current
The references inside each skill are copies of these docs pages (each names its live URL in its first lines), made again whenever a page changes, and the repository is published again on every change. Everything else in the skills comes from recorded runs against the live service, kept in the repository's runs/ folder with the script that repeats them.
Not yet
Listings in the Claude plugin directory (submitted, in review) and in the OpenAI (ChatGPT, Codex) plugin directory.