Claude Code AGENTS.md without CLAUDE.md: Sume skills install

Yes. If a repository has an AGENTS.md and no CLAUDE.md or CLAUDE.local.md in the working directory or above it, Claude Code reads the AGENTS.md as project instructions. That covers instructions only. Sume's packaged skill is installed separately with sume skills install, which writes into .agents/skills or .claude/skills.

Claude Code behavior is from Anthropic's memory page; skill and CLI behavior from Sume's Agent skills page, both read 2026-09-30.

When does Claude Code read AGENTS.md?

The memory page gives a table by file combination. With only an AGENTS.md, Claude reads it. With both AGENTS.md and a CLAUDE.md, it reads the CLAUDE.md files only. If your CLAUDE.md imports AGENTS.md, the import pulls it in. The page also says direct reading needs Claude Code v2.1.277 or later, and that some sessions cannot read it, in which case you import it from a CLAUDE.md.

Where does sume skills install put files?

The Sume docs say sume skills install writes into .agents/skills or .claude/skills. They do not say which one is chosen in a given repo, so after installing, look at which directory appeared. The same page lists sume skills list, update, export sume and remove sume; use export to review the source files before a custom install.

sume skills list
sume skills install
ls -d .agents/skills .claude/skills 2>/dev/null

Does AGENTS.md change what Sume can spend?

No. Instructions in AGENTS.md are text for the model. The CLI gate is a flag: --confirm-paid is required for generation that can reserve or spend credits. Put the rule in your instructions if you want the agent to ask first, but the flag is what the CLI checks.

Should I use MCP instead of skills?

They are separate paths. For Claude Code the hosted MCP server is added with claude mcp add --transport http sume https://mcp.sume.com/mcp, per the MCP quickstart. See allowing MCP tools in Claude Code for permission rules.