First Edition · 2026
A programmable shell for AI work. Compose engineering, security, content, and deployment workflows the way Unix composed software, one skill, one pipe at a time.
The pipe operator >> is the grammar. Skills are the
lexicon. You compose; the runtime executes.
/persona generate >> ghost "launch post" >> publish notion 60 seconds
60-second tour: one pipe builds a feature.
Video slot
Demo video lands here.
Until then, see the animated showcase for the four hero commands, or try the playground.
Recipes · One line, one outcome.
Forget the lexicon. Here are six pipes that solve real engineering problems. Copy, paste, ship. The grammar comes later.
Vulnerabilities surfaced, fixes drafted, PR opened.
luna pipe "sec >> ll-no-bluf >> pr 'security: harden auth'"scan: 12 findings (4 critical, 3 high, 5 medium) fixes: drafted for 9, manual review needed for 3 audit: all claims verified against code PR: #128 opened
Every changed file reviewed, comments inline, no LGTM bluffs.
luna pipe "rev >> ll-no-bluf --mode report-only"reviewed: 34 files, 1,204 lines comments: 17 actionable, 0 nit-picks suppressed bluffs: 0 (every recommendation cites a file:line)
Built, tested, gated, shipped to the edge.
luna pipe "test >> zen >> ship cloudflare --project luna-agent"test: 412/412 pass zen: 8/8 gates green (a11y, perf, sec, hig…) ship: deployed to luna-agent.pages.dev, custom domain live
Real docs from real code. Synced to README + Notion.
luna pipe "docs >> ll-readme-sync >> publish notion"docs: 72 functions, 12 modules documented README: counters + command index regenerated notion: page created at /docs/v2
Tests run, failures diagnosed, fixes applied, re-run until green.
luna pipe "heal --target tests"iter 1: 7 failures iter 2: 2 failures (5 fixed) iter 3: 0 failures (green) elapsed: 11m04s
Notes from git history, audited, published to GitHub + Notion.
luna pipe "ghost 'release notes' >> ll-no-bluf --mode auto-fix >> pr release"draft: 900-word notes from 23 commits audit: 4 bluffs auto-rewritten with source release: v2.1.0 published
Sample output is illustrative; real output varies by codebase. The pipe expressions are real and runnable today after install.
More recipes in the lexicon. Try one now in the in-browser playground — no install needed.
Starter ten · Forget the other 275 for now.
The lexicon has 285 entries. Most you'll meet later. These ten cover the day-one workflows — review, test, ship, audit, document, search — and compose into every recipe above.
/rev Code review with file:line citations.
full entry →/test Generate + run the test suite.
full entry →/ship Deploy to your target (Cloudflare, Vercel, npm).
full entry →/pr Open a pull request from current branch.
full entry →/sec OWASP-grade security scan + autofix.
full entry →/docs Generate docs from code, not from prose.
full entry →/rag Semantic search across your codebase.
full entry →/ghost Long-form content drafted from your code or commits.
full entry →/no-bluf Verifies every AI claim. Removes the lies.
full entry →/zen One command, every gate green: a11y, sec, perf, tests, deploy.
full entry →In the shell · Or in Claude Code · Same grammar.
The pipe operator >> composes verbs the way Unix
pipes composed software. Each verb is one job; the runtime threads
output to input. Type it in your terminal, paste it into Claude
Code, run it from CI — same grammar everywhere.
luna pipe "req >> plan >> go >> ll-no-bluf >> pr 'feat: auth'" One line does the whole sprint: read the codebase, propose tasks, implement them, audit your own claims, open the PR. No context switching.
luna pipe "ghost 'launch post' >> ll-no-bluf --mode auto-fix >> publish notion" Content lies more than code. The pipe forces every claim through a verifier before it ships. Marketing without the bluffing.
luna pipe "swarm >> heal >> zen >> ship cloudflare" Hand the build to the agents. They test, screenshot, detect, fix — on loop — until everything's green. You come back to a passing CI.
luna pipe "ll-no-bluf >> ll-readme-sync >> pr 'docs: honest'"
The luna binary is on your PATH after
npm install -g luna-agents && luna-setup.
Quote the pipe; the parser is whitespace-tolerant.
/luna-agents:pipe ll-no-bluf >> ll-readme-sync >> pr "docs: honest"
The /luna-agents:pipe command takes the same pipe
expression. Quoting rules match POSIX shell. Stages stream their
output back as the run progresses.
/ll-no-bluf and
every claim downstream is verified against the code. The runtime
refuses to ship lies.
Read the grammar, browse the lexicon, or jump straight to the install.
Configuration files describe state. Pipelines describe motion. Luna Pipes treats agent skills as words and composition as grammar: a tradition older than JSON, sharper than YAML, and built for the way work actually runs.
Forth, Joy, Factor, and Unix proved you can build a programming
language from one good idea: concatenation. f g h
means do f, then g, then h,
passing the value along. Luna Pipes adapts this for the agent era.
The unit of composition is not a function call; it is a skill
invocation. The runtime is your machine.
/persona generate >> ghost "post" >> publish notion /persona generate >> ghost "post" *4 >> publish notion /persona generate >> ? premium >> ghost "longform" : ghost "short" Most AI workflow tools express composition as configuration. Luna Pipes expresses composition as code, the same trick Unix played in 1973. Side by side, on the same job, four personas writing a launch post each, then posting to Notion.
steps:
- action: persona_generate
- action: ghost_post
params:
title: "launch post"
repeat: 4
- action: publish_notionpersona generate >> ghost "launch post" * 4 >> publish notionMany of the lexicon entries call one of the services below. Each is accessed through a verb in the lexicon, not a bespoke integration page. The runtime carries credentials; the pipe carries intent.
npm install -g luna-agents Luna Agents ships as a CLI and a Claude Code plugin. Both expose the same lexicon. Pipes work in your terminal, your editor, and your CI. See Chapter 04 for platform-specific notes.