agentfiles

My Claude Code and Codex setup — memory, subagents, skills, plugins — as a repo you clone and link. Mac and Linux, one script, no sudo.

Install

git clone git@github.com:xs1128/agentfiles.git ~/.agentfiles
sh ~/.agentfiles/install.sh
exec $SHELL

Installs Claude Code, Codex, bun and rtk if missing, symlinks claude/ into ~/.claude and codex/ into ~/.codex, and installs the pinned plugins. Needs curl, git, jq.

Rerun it to update. Safe to rerun.

flag	effect
--no-deps	skip the installs
--no-bootstrap	skip the plugins
--no-codex	Claude Code only
--mcp	also install the MCP servers (needs node)

Because both directories are symlinks into the clone, editing a skill here takes effect next session, in either harness.

Codex

Three things, nothing else: rtk, the skills, the bundled plugins.

codex/AGENTS.md   → ~/.codex/AGENTS.md   the rtk rule
skills/*          → ~/.codex/skills/*    the same 45 directories Claude gets
plugins.toml.tmpl → appended to ~/.codex/config.toml

Codex has no PreToolUse hook, so rtk is an instruction in AGENTS.md rather than a rewrite. It costs a few tokens per session that the Claude side gets for free.

config.toml is Codex's own — it writes to it during sessions — so the plugin enables are appended instead of symlinked, and only for a block the file does not already have. Nothing there is ever rewritten, so a plugin you turned off by hand stays off, and the previous copy lands in ~/.codex/backups/. Model and reasoning effort are deliberately not pinned: they are set in Codex's UI, and writing them here would revert that on every install.

Re-link without touching the rest:

sh scripts/link-codex.sh

GLM

Put the key in ~/.agent.env, readable only by you:

install -m 600 /dev/null ~/.agent.env
echo 'ZAI_API_KEY=your-key-here' >> ~/.agent.env

Then:

glm

Same config against z.ai: opus → glm-5.2, sonnet → glm-5-turbo, haiku → glm-4.7. Plain claude is unaffected. ZAI_API_KEY in the environment wins over the file.

MCP

Not installed by default, and never loaded implicitly — a server costs tool definitions in every prompt.

sh scripts/bootstrap.sh --mcp

Then load per session:

claude --mcp-config ~/.claude/mcp/web.json --strict-mcp-config    # playwright
claude --mcp-config ~/.claude/mcp/cloud.json --strict-mcp-config  # firebase

Playwright's tools error until a browser is installed: npx playwright install chromium.

Skills

Loaded by both harnesses out of skills/. Vendored from upstream at pinned commits. To pull newer ones:

sh scripts/sync-skills.sh

That recopies from the commits manifests/skills.json names, so upstream changes arrive as a reviewable diff.

Plugins

caveman and claude-hud only, pinned in manifests/plugins.json. Change a pin and rerun sh scripts/bootstrap.sh.

Layout

claude/       symlinked into ~/.claude
codex/        symlinked into ~/.codex
skills/       loaded by both
manifests/    pinned plugins, MCP packages, skill sources
scripts/      link, link-codex, bootstrap, glm, sync-skills
install.sh

Uninstall

Everything this repo puts on the machine is a symlink into the clone, so unlinking is most of the job:

find ~/.claude ~/.codex -maxdepth 2 -type l \
  -exec sh -c 'readlink "$1" | grep -q "/.agentfiles/" && rm "$1"' _ {} \;
rm -f ~/.local/bin/glm ~/.codex/skills/.agentfiles-managed

Then drop the # agentfiles block from your shell rc. Whatever was replaced is under ~/.claude/backups/ and ~/.codex/backups/. The plugin blocks appended to ~/.codex/config.toml are left behind; delete them by hand.

The tools are a separate question, and both of these throw away auth and history:

rm -rf ~/.local/bin/claude ~/.local/share/claude ~/.claude ~/.claude.json
rm -rf ~/.codex ~/.local/bin/codex ~/.local/bin/codex-code-mode-host

Codex keeps its binary under ~/.codex, so that one line takes the install with it. bun is rm -rf ~/.bun. rtk ships no uninstaller — delete whatever which rtk names.

Attribution

Most skills are other people's work, MIT licensed and vendored. See CREDITS.md.