Agent skills

Two agent skills ship with the nace CLI for code agents such as Claude Code, Codex, Cursor and Pi:

You create and fill a workspace once; after that every question is a search. Keeping the two apart keeps the search skill to the navigate tools.

Install the local MCP server with them. The skills run only nace commands, but the agent works better with the server's tools beside them: it can upload a local file, check a job, run automatic_search or qa when you want Nace's search agent to answer in one call, and parse, extract from or classify a document without leaving the session. The setup below adds both, and the server signs in with the key nace login saves.

nace skill install writes the skills from the CLI you have installed, so they always match its commands. After you upgrade the CLI, run it again: nace locate, nace read-file, nace run-sql, nace folder-metadata and nace catalog print a one-line note when an installed copy is older than the CLI.

Install the skillsCopy Anchor Link

Install the CLI, sign in, add the local MCP server, and write both skills to your agent's skills folder:

uv tool install nace-cli                 # or: pipx install nace-cli
nace login                               # paste the nace_sk_ key when it asks
claude mcp add nace -- uvx nace-mcp     # the local MCP server
nace skill install --agent claude-code
# wrote /Users/you/.claude/skills/drex-search/SKILL.md
# wrote /Users/you/.claude/skills/drex-workspace/SKILL.md

Then start a new session of the agent: skills load when a session starts.

nace skill install writes each skill as <name>/SKILL.md; name one to write only it, such as nace skill install drex-search --agent claude-code. Running it again updates them. Add --project to write to the project's folder, under the current directory, instead of your user one.

The agent runs nace in its shell, so the CLI must be on its PATH and the key saved with nace login: a NACE_API_KEY set only in the local MCP server's config doesn't reach the agent's shell. drex-search needs a workspace to search: set one up with the drex-workspace skill, the MCP server's workspace tools, or the dashboard, and save it as the default with nace workspace-use ID (or the MCP server's use_workspace). An account with only one workspace needs neither.

Use themCopy Anchor Link

Set the workspace up once, such as "create a workspace from the files in ./contracts and save it": the agent follows drex-workspace. Then ask questions about the documents, such as "what is the monthly base rent for unit 12 in the first lease year?": the agent follows drex-search, in this session or a later one.

drex-search tells the agent to:

1. Locate what the question needs: nace locate "<need>" --question "<the question>".
2. Read the pages it names with nace read-file, and query the tables it names with nace run-sql.
3. Locate again for a part still missing, passing the question's locate id: nace leaves out what the agent was already shown and read, so each locate shows something new. When the answer needs every file of a kind, such as a total over each invoice, nace locate --all lists the members to read.
4. Stop when every fact comes from a page it read or a query it ran, and cite each one with its file and page.

It also says which command reads which kind of file (spreadsheets and CSV files with nace run-sql, documents with nace read-file), the SQL rules the document service's own search agent follows, and to report something missing only after two locates have not found it.

nace locate, nace read-file, nace run-sql, nace folder-metadata and nace catalog are free. Locates are limited to 150 per account every 10 minutes, and file reads, SQL queries, folder listings and catalog queries together to 600: about 30 questions every 10 minutes. Ingestion is a workspace job.