# Browser — zero install. Live now.
open https://ruvnet.github.io/agent-harness-generator/
# OR — terminal scaffold
npx create-agent-harness my-bot --template vertical:coding --host claude-code
cd my-bot && npx . --help

That's it. You now own an npm-publishable AI agent harness with your name on it, your branding, your agents, your MCP server, your memory, and an npx <your-name> command. It runs on Claude Code, Codex, pi.dev, Hermes, OpenClaw, or RVM — pick one or all. Nothing leaves your browser or your machine.

Open the Studio

One line: A meta-harness — a marketplace plugin + CLI that scaffolds your own focused, vertical AI agent harnesses with their own npx <name> command, MCP server, memory, learning loop, and brand — that run unchanged on Claude Code, Codex, pi.dev, Hermes, OpenClaw, and RVM.

What's a meta-harness? A harness is a runtime that orchestrates AI agents (memory + routing + hooks + MCP + claims). A meta-harness is the level above: a harness whose product is OTHER harnesses. agent-harness-generator emits self-contained, npm-publishable harnesses you OWN — same kernel, your branding, your agents, your marketplace presence. The kernel updates flow to your harness via @ruflo/kernel npm peer; the content stays yours.

One paragraph: Ruflo bundles primitives (MCP server, hooks, memory bridge, swarm coordinator, intelligence pipeline, claims, routing) WITH opinionated content (60+ agents, 30+ skills, 33 plugins). agent-harness-generator factors those apart. You pick the primitives, pick the content, supply a name + brand, and out comes a brand-new npm-publishable harness with its own CLI, MCP registration, memory namespace, and marketplace identity — running on the host of your choice.

A 100% client-side Studio (in the spirit of ruflo's goal UI) that turns any GitHub repo — or a blank slate — into a governed, branded, multi-host AI agent harness. Recommend agents, skills, commands, MCP tools, and policy; preview the live file tree; download a signed-ready, npm-publishable .zip. Nothing leaves your browser. Desktop- and mobile-friendly, deployable to GitHub Pages.

Open the Studio design ADR-020→024

Agent Harness Studio

Embeddings recommend · rules generate · tests prove parity.

Repo → Harness Author a Claude skill/agent/command In-browser Verify

Mobile-friendly

MCP is included as a first-class adapter surface, not the core identity. It is modular, gated, and default-deny (ADR-022):

• Modes: off · local (stdio) · remote (Streamable HTTP + auth).
• Emits src/mcp/{server,tools,resources,prompts,policy,audit}.ts (+ auth.ts remote) and a scannable .harness/mcp-policy.json.
• Safe defaults: default-deny, no network/shell/file-write, approve-dangerous, 30 s timeout, 8 calls/turn, audit on.
• harness mcp-scan <path> — "npm audit for agent tools": static-only scan (never executes) flagging shell/network grants, missing audit/timeouts, wildcard permissions, unguarded secrets, and unpinned deps. Exit 1 on any HIGH.

CLI Repo → Harness (ADR-026) — the deeper, local counterpart to the Studio's importer:

harness analyze-repo . # local, analysis-only → repo-profile.json + harness-plan.json
harness analyze-repo . --embed # opt-in deterministic embeddings via @ruvector/ruvllm (offline; lexical fallback)
harness analyze-repo . --scaffold my-harness # materialise the recommended harness

No repository code is executed; inferred build/test commands are emitted as trust: inferred · execution: disabled.

cd apps/web-ui
npm install
npm run dev # local Studio
npm test # 48 generator unit tests
npm run e2e # Playwright desktop + mobile (zero console errors)
npm run bench # generator hot-path micro-bench (sub-100μs/op)

Source + rationale: apps/web-ui/ · ADRs 020 · 021 · 022 · 023 · 024 · 025.

One canonical catalog (templates/catalog.def.mjs) feeds the CLI templates, the catalog.json the template-catalog Rust crate validates, and the Studio gallery — so UI, CLI, and core never drift.

npx create-agent-harness --list # browse the catalog
npx create-agent-harness my-bot --template vertical:coding

Each generated vertical ships bespoke domain agents (with system prompts), skills, commands, and guarded per-host settings. The 11 generated dirs come from the definition; minimal, devops, trading, support, legal, and research are curated by hand and never clobbered by the generator.

Latest vertical — vertical:education (iter 80): a tutoring pod (tutor → explainer → quiz-master → grader) over per-learner mastery memory. Pedagogy invariants are policy: abstain-not-hallucinate floor, no teaching on unmastered prereqs, hidden rubrics. Two commands: teach-next (run a teaching cycle) and mastery-report (summarise the learner's map). MCP namespaces mastery_log + curriculum; everything else default-deny.

# 50ms scaffold → validate → cleanup, no network, exits 0 if healthy
node examples/quickstart/quickstart.mjs
# Try every supported host (claude-code | codex | pi-dev | hermes | openclaw | rvm)
node examples/quickstart/quickstart.mjs --host=codex
# 20ms two-instance federation handshake demo
node examples/federation/federation.mjs

See examples/ and docs/ARCHITECTURE.md for the layered map.

Production-ready release pipeline. Full CI matrix green: 16 jobs across Rust ×ばつ 3 OS + WASM ×ばつ 3 OS + Node 20+22 ×ばつ 3 OS + Bench + pack+install ×ばつ 3 OS + CI-passed aggregator. The release flow is a single command (node scripts/release.mjs <bump> --push) that bumps 15 sources atomically, runs all gates, and tags. See ADR-019 for the architectural lockdown.

 You <- harness AUTHOR (uses the meta-harness)
 |
 v
 agent-harness-generator <- THE META-HARNESS
 (this repo: scaffolds, signs, publishes harnesses)
 |
 v
 Your harness (npm package) <- THE HARNESS YOU SHIP
 |
 v
 Your users
 |
 v
 npx <your-name> <- Identity (rename + brand)
 |
 v
 <your-harness> <- Content (your agents/skills/plugins/prompts)
 |
 v
 @ruflo/kernel <- Kernel (shared primitives, Rust + WASM + NAPI-RS)
 |
 v
 Host adapter <- Per-host abstraction
 (Claude Code / Codex / pi.dev / Hermes / OpenClaw / RVM)
 |
 v
 LLM providers

Read top-down: you (the harness author) operate the meta-harness. The meta-harness produces your harness. Your harness is what users install. They never see the meta-harness layer — only the brand and CLI you ship.

The kernel is Rust source code compiled to two targets: WebAssembly (primary, cross-platform) and per-platform native binaries via NAPI-RS (escape hatch for hot Node paths). At load time, @ruflo/kernel prefers the native package for the current platform and falls back to wasm.

Working precedent: @ruvector/emergent-time@0.1.0 — 55 KB wasm-opt'd module shipping today through exactly this Rust → wasm-pack → npm pipeline.

See ADR-004 — Host integration model.

git clone https://github.com/ruvnet/agent-harness-generator
cd agent-harness-generator
# Rust workspace
cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
# WASM build
npm run build:wasm
# TypeScript + smoke
npm install
npm run build
npm run smoke

See CONTRIBUTING.md for the full developer workflow.

Start here: docs/adrs/INDEX.md

Highlights:

• ADR-001 Goals & non-goals — what this is and isn't
• ADR-002 Kernel boundary — Rust + WASM + NAPI-RS
• ADR-002a Publishing pipeline — Cargo workspace + wasm-pack + napi build
• ADR-003 Generator architecture — create-vite-style templates + AST-aware rename
• ADR-004 Host integration — adapter contract per host
• ADR-006 Memory + learning — @ruvector/emergent-time@0.1.0 integration
• ADR-009 Anti-slop — derived trust tiers
• ADR-011 Witness + provenance — signed manifests

• ruflo — the meta-harness this generator factors apart
• ruvector — vector + agentic database (memory backend)
• @ruvector/emergent-time — memory-decay clock the kernel uses
• NAPI-RS — Rust → Node bindings used for the native target

MIT — see LICENSE.

Keywords: agent harness, agent harness generator, AI agent scaffolding, MCP server, Claude Code plugin, Codex plugin, pi.dev extension, hermes agent, multi-agent framework, agentic AI, agentic workflow, autonomous agents, agent orchestration, vertical AI harness, agent CLI generator, npm create agent, npx scaffold, Rust WASM kernel, NAPI-RS, wasm-bindgen, wasm-pack, agent memory, ReasoningBank, HNSW, emergent time, agent provenance, witness manifest, plugin marketplace, IPFS registry, drift detection, anti-slop, TDD, self-evolving agent, federated agents, swarm intelligence, GCP Workload Identity Federation, Secret Manager, SLSA provenance, npm provenance