ner-jarvis

One command to configure Claude Code for Northeastern Electric Racing. ner-jarvis installs NER's onboarding skills and connects its data sources (Atlassian, GitHub, Slack) into your local Claude Code — so a new software-team member goes from a fresh ~/.claude to a working setup in a single step.

A CLI scoped to App Software, made to accelerate onboarding and surface shared development bugs quickly. The skills it ships live in this repo's skills/; the CLI embeds them at build time.

What it installs

Into ~/.claude/skills/, six NER skills:

skill	what it does
ner-onboard	walks a new member through getting started
ner-ask	answers NER software questions from Confluence + GitHub
ner-setup	environment / dev-setup help
ner-repo-explainer	explains what a given NER repo does
ner-escalation-router	points you at the right owner / channel
ner-flag-stale	records a doc that misled you, for ner-jarvis stale

And two data sources (both plugins):

source	mechanism
slack	plugin (slack@claude-plugins-official)
atlassian	plugin (atlassian@claude-plugins-official)

GitHub isn't a connected source — the skills use the gh CLI (Claude Code has a shell, and gh is more capable for READMEs, commits, CODEOWNERS, and search). ner-jarvis checks gh and guides gh auth login; it never installs or authenticates it for you.

Everything ner-jarvis adds is tracked in ~/.claude/ner-jarvis/state.json, so it only ever touches its own additions — it never clobbers skills or sources you added yourself. State carries a migrationLevel alongside stateSchemaVersion: every persisted shape is versioned, and any structural change is migration-governed (a schema-snapshot guard test fails CI otherwise). Two ladders handle upgrades — pure schema migrations applied when state is read, and side-effecting environment migrations run only during setup/update. doctor (like every read) is read-only and never relocates or rewrites your files. Every setup/update also writes a per-run undo record (under ~/.claude/ner-jarvis/undo/) capturing what it installed and the prior content of anything it overwrote, so ner-jarvis undo can reverse a run precisely — and rollback is always performed by the version of ner-jarvis that made the run (the current binary in-process, or an archived/pinned older binary via delegation), so the tool can evolve without carrying a back-compatibility tax.

Install

No GitHub account, no org membership, and no gh needed — the repo and its release assets are public.

macOS / Linux:

case "$(uname -s)" in Darwin) os=darwin;; Linux) os=linux;; *) echo "unsupported"; exit 1;; esac
case "$(uname -m)" in arm64|aarch64) arch=arm64;; x86_64|amd64) arch=x64;; *) echo "unsupported"; exit 1;; esac
curl -fsSL -o ner-jarvis \
  "https://github.com/Northeastern-Electric-Racing/ner-jarvis-cli/releases/download/v0.1.0/ner-jarvis-bun-$os-$arch"
chmod +x ner-jarvis && ./ner-jarvis

To keep it on your PATH instead of the current folder, swap the last line for mkdir -p ~/.local/bin && mv ner-jarvis ~/.local/bin/ && ner-jarvis.

Windows (PowerShell):

Invoke-WebRequest -OutFile ner-jarvis.exe `
  "https://github.com/Northeastern-Electric-Racing/ner-jarvis-cli/releases/download/v0.1.0/ner-jarvis-bun-windows-x64.exe"
.\ner-jarvis.exe

Running it with no arguments starts the interactive setup. It prompts before each step, and ner-jarvis undo reverses a run precisely.

npx (Node ≥ 18) — not published yet, see below:

npx ner-jarvis@latest

Platforms

Platform	Asset
macOS (Apple Silicon)	ner-jarvis-bun-darwin-arm64
macOS (Intel)	ner-jarvis-bun-darwin-x64
Linux x64	ner-jarvis-bun-linux-x64
Linux arm64	ner-jarvis-bun-linux-arm64
Windows x64	ner-jarvis-bun-windows-x64.exe

Each is a standalone binary with the skills payload compiled in — no runtime, no node_modules, no network fetch at install time.

Still to do: npm

npx ner-jarvis@latest is the one install path that doesn't work yet. It needs the NPM_TOKEN repository secret set; until then the release's npm job builds the bundle, reports the skip, and stays green. Setting it and re-running that job is all that's required — the binaries above are unaffected either way.

Build from source

Build it yourself with Bun:

bun install
bun run build:binary   # → dist/ner-jarvis   (standalone binary for this host)
# or
bun run build:npm      # → dist/index.js     (Node-runnable bundle)

Both build steps embed the skills + sources payload first, so the output is self-contained.

Commands

ner-jarvis                   # setup: install all skills + connect all sources (default)
ner-jarvis setup slack       # targeted: act only on the named skill/source
ner-jarvis update            # re-apply the latest embedded payload
ner-jarvis doctor            # read-only health check (non-zero exit if anything's off)
ner-jarvis uninstall         # remove only what ner-jarvis installed
ner-jarvis undo              # reverse the most recent setup/update run
ner-jarvis undo --list       # show the recorded run stack

Useful flags:

ner-jarvis --dry-run         # preview every change; write nothing
ner-jarvis --force           # overwrite items we installed but you've since edited
ner-jarvis --yes, -y         # run unattended (auto-confirm every step)

• setup is interactive on a TTY (confirm each step: Enter = yes, n = skip), idempotent, and non-destructive — safe to re-run. A full run also clones the NER workspace into a directory you choose and opens Claude Code there.
• update reuses the same convergence engine as setup; on the binary channel it reports a newer release but never rewrites the running executable.
• uninstall is interactive on a TTY too: it confirms each removal — the skills, each source (so you can keep a plugin you use outside NER), and the global note — with Enter = remove, n = keep (--yes removes everything unattended). It only ever removes ner-jarvis's own additions and never removes shared marketplaces.
• undo reverses a recorded run rather than everything ner-jarvis tracks: with no flags it undoes the most recent run, interactive on a TTY (Enter = undo, n = keep; --yes reverses everything unattended). It hash-guards edited skills (left in place unless --force), restores the prior content of anything the run overwrote, never removes shared marketplaces, and marks a run done only when fully reversed. undo --list prints the recorded run stack. A run made by a different ner-jarvis version is delegated to the binary that made it — an archived per-version binary, or npx ner-jarvis@<version> — so the current binary never parses an old version's records. (A full uninstall clears the undo journal and those archived binaries; the decision log is kept.)
• name… (e.g. slack, github) can be passed to setup / update / doctor to scope the run; a targeted run never removes anything.

Authentication

ner-jarvis never handles your secrets. It connects the sources but cannot perform OAuth for you. After setup, authenticate each yourself:

• Slack, Atlassian (plugins) — prompt on first use inside Claude Code.
• GitHub — install the gh CLI and run gh auth login.
• Any mcp source — claude mcp login <name>, or open Claude Code and run /mcp.

Troubleshooting

Run the health check to see what's missing or still needs auth:

ner-jarvis doctor
ner-jarvis doctor github     # scope the check to one source

doctor is read-only and exits non-zero if any tracked skill or source is missing or unhealthy.

Development

bun install
bun run embed    # regenerate the embedded payload from skills/ + sources.json
bun test

• The NER skills live at skills/ and are embedded at build time (via bun run embed) — there's no runtime fetch.
• Data sources are declared in sources.json.
• behavior.md is the language-agnostic contract — the authoritative spec for command surface, wizard steps, state shape, reconciliation rules, and exit codes. Read it before porting ner-jarvis to another runtime.