NoizuPromptLingo (NPL)
Repo: https://github.com/noizu-labs-ml/NoizuPromptLingo
The NPL structured prompt-engineering syntax — placeholders, intuition pumps, directives, prefixes, and tagged prompt sections — plus a public MCP that serves that corpus. This repo publishes the language agents load; the agent work platform (sessions, tickets, chat, wiki, memory, tobor MCP fleet) lives separately in agent-kit-mcp.
What
- Convention YAML in
conventions/ — single source of truth for the NPL syntax sections (syntax, declarations, pumps, directives, prefixes, prompt-sections, special-sections, fences).
- MCP at
/mcp — two tools only: NPLLoad (expression DSL) and NPLSpec (versioned spec generation). No authentication required.
- OAuth is optional. Clients that insist on OAuth (Claude Code, Cursor, …) complete PKCE and receive a site-approval token (
sub=site:npl). The token does not identify a user; it only means "approved for this site."
- REST at
/api/v1/npl/* (public) for browsing sections and generating the same spec the admin panel produces.
- Admin / conventions browser (signed-in) to inspect components and generate NPLSpec blocks.
Hook it up:
claude mcp add --transport http npl https://promptlingo.dev/mcp
Browse as files (VFS — mount the syntax as markdown, no token):
mcp-mount --url wss://promptlingo.dev/vfs --mount ~/npl --ro
# then open ~/npl/tobor/_npl/sections/syntax.md
Kernel FUSE (mcp-fuse) is for a local unix-socket VFS. Setup, binaries (macOS arm, Linux amd64/arm64, Windows amd64/arm64), and WinFsp notes: docs/howto/vfs.md.
Why
Prompt conventions scattered across ad-hoc markdown drift and can't be loaded selectively. NPL keeps one versioned YAML source and serves it on demand: agents pull exactly the syntax sections they need (NPLLoad(expression="syntax#placeholder:+2 directives -pumps")), editors curate the corpus through an admin UI, and every consumer sees the same spec. It is the shared vocabulary underneath the Noizu agent fleet.
Getting Started
Prerequisites: Docker + Docker Compose; Make.
make init # Generate .env files with secrets
make build # Build backend + frontend + nginx images
make run # Start the stack (nginx on :8080)
Hot-reload dev mode (source-mounted, Next.js HMR + Phoenix code reloading):
make run-dev # Foreground with live logs
make run-dev-d # Detached
make logs-dev # Tail logs
make stop-dev # Tear down
A live-sandbox mode (single container, Node + Elixir + Samba, mountable as a network share for remote editing) is also available — see make sandbox / make run-sandbox / make sandbox-mount.
Migrations (Liquibase for canonical schema, Ecto for app-level):
make migrate
make migrate-status
make migrate-rollback
Deployment is CI-driven to Kubernetes via the Helm chart in helm/ (registry ops.noizu.com, secrets via Infisical).
How It Works
- Three-container stack — Next.js 15 frontend, Phoenix 1.8 (Elixir) API backend, nginx reverse proxy — sharing Postgres (PostGIS, pgvector) and Redis. Schema is Liquibase-canonical + Ecto app-level; Weaviate powers semantic MCP tool search (intent mode).
conventions/*.yaml is rendered on request: NPLLoad for quick agent-facing snippets via expression DSL, NPLSpec for full ⌜[email protected]⌝-wrapped spec blocks with dependency resolution. npl/npl-full.md is a generated artifact of that pipeline.- Human sign-in (Authentik OIDC) is only for the conventions admin UI; the MCP endpoint itself is open.
Repo Layout
NoizuPromptLingo/
├── conventions/ # YAML source of truth for NPL syntax sections
├── frontend/ # Next.js 15 — public landing + conventions admin
├── backend/ # Phoenix 1.8 — NPLLoad/NPLSpec MCP at /mcp, REST /api/v1/npl
├── src/ # Python npl_mcp tooling (npl-docs-regen, local tooling)
├── nginx/ # Reverse proxy
├── helm/ # Kubernetes Helm chart
├── liquibase/ # Canonical schema changelogs
├── design/ # Theme YAML (npl-brutalist, npl-editorial, npl-minimal, npl-nocturne)
├── docs/ # Architecture and convention pipeline (PROJ-ARCH / LAYOUT / SCHEMA + summaries)
└── Makefile # Build/run/deploy targets
project-management/ holds UX-research personas, user stories, screens, and components generated as design/planning input — distinct from the product's own "Agent Persona" and "Customer Persona" domain entities.
License: see LICENSE.
