| name | cortex-setup |
| description | Bootstrap Cortex on a fresh machine. Trigger this skill whenever the user wants to install, set up, configure, or first-time-use Cortex / the `cortex` CLI / the Cortex memory atlas — including phrases like 'install cortex', 'set up cortex', 'cortex non funziona', 'first time using cortex', 'configura cortex', 'help me get cortex working', or when the user says they cloned the repo and want to start. Also use this when `cortex` is missing, when `~/cortex/` doesn't exist, or when the user wants to point Cortex at a remote, switch image provider, or start the React atlas. Do NOT use this for daily memory work — that's the `cortex` skill. |
Cortex Setup
This skill brings a fresh machine to a working Cortex install:
- the
cortex CLI on $PATH (Python, via pipx),
- the Cortex agent skills (
cortex, cortex-setup, cortex-scan) installed user-level at ~/.claude/skills/ so they're available in every Claude session, not only inside $CORTEX_SRC,
- an initialized memory store at
~/cortex/,
- a chosen image provider (codex default → OpenAI fallback),
- an optional GitHub remote,
- the React atlas running locally (optional).
Layout
The user's data lives at ~/cortex/ (configurable via $CORTEX_ROOT). The Cortex source repo can live anywhere the user prefers — call this path $CORTEX_SRC. The repo contains:
cortex-cli/ — Python CLI source (installed globally via pipx)cortex-fe/ — React atlas front-end.claude/skills/ — the cortex, cortex-setup, and cortex-scan agent skills. During setup these are symlinked into ~/.claude/skills/ so they're available in any Claude session (not only when inside $CORTEX_SRC). Symlinks (not copies) keep them in sync with git pulls in the source repo.
When you run setup, the very first thing to nail down is $CORTEX_SRC. Either ask the user where they have the repo, or — if they don't have it yet — agree on a path and clone it there.
How this skill behaves
You are not an installer wizard — you are a colleague helping the user finish a setup. Read what's already in place before running anything destructive. If the user already has cortex on PATH and ~/cortex/ exists, do NOT re-run cortex init — instead, summarize the state and ask what's missing.
Before each step, check first → act second. Tell the user what you're about to do in one short sentence. Don't narrate every command.
The setup, step by step
1. Verify Python and pipx
python3 --version
pipx --version 2>/dev/null || python3 -m pip install --user pipx && python3 -m pipx ensurepath
If pipx had to be installed, the user must restart the shell (or source ~/.zshrc) before pipx commands work. Tell them.
2. Locate or clone the Cortex source repo
Pin down $CORTEX_SRC — the path to the cortex source repo (containing cortex-cli/, cortex-fe/, .claude/skills/).
A. Already cloned: ask the user where it is, or look for likely locations (e.g. recent cd history, common ~/code / ~/dev / ~/Documents parents). Confirm with ls $CORTEX_SRC/cortex-cli/pyproject.toml.
B. Not yet cloned: ask the user where they want it. Common picks: ~/code/cortex, ~/dev/cortex, or wherever they keep other projects. Then:
git clone <repo-url> $CORTEX_SRC
You need the repo URL from the user — don't guess it.
3. Install the cortex CLI globally
pipx install -e $CORTEX_SRC/cortex-cli
pipx inject cortex-cli Pillow
Verify: cortex --version should print a number.
If cortex: command not found afterwards, the user needs to restart the shell or source ~/.zshrc (pipx puts it in ~/.local/bin and only adds that to PATH on first install).
4. Install the agent skills user-level
The three skills (cortex, cortex-setup, cortex-scan) live at $CORTEX_SRC/.claude/skills/. To make them available in every Claude session — not only when working inside $CORTEX_SRC — symlink each into ~/.claude/skills/:
mkdir -p ~/.claude/skills
for s in cortex cortex-setup cortex-scan; do
ln -sfn "$CORTEX_SRC/.claude/skills/$s" "$HOME/.claude/skills/$s"
done
Symlinks (not copies) keep the user-level skills in sync with git pulls in the source repo automatically.
Quick check — all three skill files should resolve:
ls ~/.claude/skills/cortex/SKILL.md ~/.claude/skills/cortex-setup/SKILL.md ~/.claude/skills/cortex-scan/SKILL.md
If any is missing in $CORTEX_SRC, the source repo is incomplete or stale — git pull (or re-clone).
5. Initialize the memory store (or restore from remote)
cortex init
cortex init is smart: it scans every authenticated gh account for a private cortex-memory repo and, if it finds exactly one, clones it into ~/cortex/ instead of starting empty — so memory restores cleanly across machines. With no remote available (or no gh), it falls back to creating an empty store.
Outcomes:
- Restored (one
cortex-memory repo found) → printed as restored: github.com/<owner>/cortex-memory. origin is already wired; step 7 can be skipped.
- Empty init (no remote / no gh) → classic blank store with
notes/, inbox/, daily/, assets/, .config/, and a scoped .git/.
- Ambiguous (multiple
cortex-memory repos) → command refuses to guess. Ask the user which account, then re-run with cortex init --from <OWNER>.
Flags worth knowing:
--no-restore — always start empty, skip the gh probe.--from <OWNER> — clone from github.com/<OWNER>/cortex-memory.git directly (skips detection).--no-git — don't create the scoped git repo.
If ~/cortex/notes/ already exists, cortex init is a no-op and tells you to run cortex stats. Don't re-run it.
6. Pick an image provider
Default is codex (uses the local codex CLI as a pure prompt → PNG generator, sandboxed to $CORTEX_ROOT). Fallback is openai (HTTP to images/generations).
which codex
If codex is not installed OR the user prefers OpenAI:
cortex config set image.provider openai
cortex config set image.openai.token sk-...
If the user has neither codex nor an OpenAI key, image generation simply won't work — the rest of the CLI still does. Note this and move on; don't block setup.
The image style preset (image.style) is pre-configured in defaults — editorial isometric, parchment palette. Don't override unless the user asks.
7. (Optional) Configure a GitHub remote
Skip this step if step 5 cloned from cortex-memory — origin is already wired, verify with cortex remote info and move on.
Otherwise, the memory store can sync to a private GitHub repo. This step requires gh.
which gh && gh auth status
If gh is missing or unauthenticated: skip this step. Tell the user the memory still works locally and they can run cortex remote create later when gh is set up. Do NOT try to install gh yourself.
If gh is available:
gh auth status
cortex remote create --user <username>
If multiple gh accounts are authenticated, the --user flag is required (the CLI auto-switches gh accounts based on the remote's owner and restores the previous account afterwards — no manual gh auth switch needed).
Verify: cortex remote info.
8. (Optional) Start the React atlas
The visual atlas at $CORTEX_SRC/cortex-fe/ reads the store via a Vite bridge plugin that shells out to the cortex CLI. The CLI ships with atlas serve, which auto-discovers the frontend path, runs npm install if needed, and starts the dev server. You — the agent — should run this automatically, not ask the user to cd and npm by hand:
cortex atlas serve
cortex atlas serve --port 3000
cortex atlas serve --host 0.0.0.0
cortex atlas serve --reinstall
How atlas serve finds the frontend:
$CORTEX_SRC env var, if set (expected to point at the source repo root containing cortex-fe/).- Otherwise, the package install path — works automatically when the CLI was installed editable (
pipx install -e <repo>/cortex-cli, which is what step 3 does).
This means after a normal setup the agent can launch the UI with one command — no cd, no path discovery, no manual npm. Run cortex atlas serve in the background (so the dev server keeps running) and tell the user the URL.
Requires npm (Node.js) on PATH. If npm is missing, the command fails with a clear message — install Node and retry, or skip if the user only wants the CLI.
If the user only wants the CLI (no UI), skip this step.
9. Sanity check
cortex stats
cortex config get
cortex --help
Tell the user what's set up, what's optional and skipped, and point them at the cortex skill (the daily-work skill) for what to do next.
Common pitfalls
cortex: command not found after pipx install: the user needs to restart the shell or source ~/.zshrc. pipx put it in ~/.local/bin and added that to PATH only on first install.ModuleNotFoundError: PIL: forgot pipx inject cortex-cli Pillow. Re-run.codex did not produce an image: the codex CLI isn't on PATH, or is on a sandboxed path. Either install codex or switch to openai (step 6).gh: multiple accounts authenticated: pass --user <name> to cortex remote create. The CLI handles the switch transparently.cortex atlas serve says "cannot locate cortex-fe/": the CLI was installed non-editable (pipx install without -e). Either reinstall with pipx install -e <repo>/cortex-cli, or set CORTEX_SRC=/path/to/cortex-repo in the environment.- Multiple
cortex-memory repos found during step 5: stop and ask the user. Cloning the wrong one mixes a different machine's memory in.
- Memory store at a non-default path: set
CORTEX_ROOT=/some/other/path in the shell environment before any cortex command. Persist in the shell rc if they want it permanent.
What you do NOT do here
- Do NOT assume where
$CORTEX_SRC is — ask the user.
- Do NOT write any notes during setup. That's the
cortex skill's job.
- Do NOT generate images during setup — there's nothing to generate yet.
- Do NOT run
cortex init if the store already exists.
- Do NOT install
codex or gh for the user — those are out of scope; just detect them and skip cleanly when missing.
- Do NOT push to remote unless the user explicitly asked for a remote.
