Every repo, its own harness
You copy .claude/ from one project to the next. Keeping it in sync is manual and error-prone.
open source · 5 engines
The multi-agent harness for any repo.
Agents, skills, hooks, permissions and memory — versioned in one config and rebuildable with a single command. Renders to Claude Code, Codex, Cursor, Copilot and AGENTS.md without overwriting a line of yours.
$
npx navori init
MIT
Node ≥ 20
0 runtime deps
Claude Code
$ navori init --recommended
✓ stack: Next.js · pnpm
✓ engines: claude · codex
✓ plugins: engram · codegraph · tgrep
→ rendering managed assets
✓ CLAUDE.md · AGENTS.md
✓ .claude/ agents · skills · hooks
✓ .claude/settings.json permisos
✓ Done — 34 created · 0 overwritten
22 commands 12 presets 7 plugins 5 engines 8 agents 41 skills
MIT Node ≥ 20 0 runtime deps
v0.9.0
Anatomy of a session
Follow one prompt through the harness
With no harness, your sentence reaches a model that decides alone. This is everything navori puts in its path — eight stops, from the sentence to the commit, and back into memory.
your prompt
fix the login bug support reported
You type a sentence. Nobody has thought yet.
Before the model reads your prompt, a startup hook has already run. That's the difference between an assistant that starts blank every morning and one that opens the session knowing where it stands.
The walkthrough runs on its own. Tap a stop to jump to it.
The problem
Every repo reinvents the same thing
If you work with multiple projects and AI agents, you know exactly what we mean.
Every engine, its own config
Claude has .claude/, Cursor has .cursor/, Copilot has .github/copilot-instructions.md. Same intent, three files.
SDD erodes over time
Skills, hooks, and Spec-Driven Development conventions decay when there's no way to version and roll them forward.
Scopes
The harness no longer lives only in the repo
Three additive scopes, each opt-in. None degrades the one below it: the machine layer steps aside the moment it finds a repo with navori.
per repo
The repo
A checked-in navori.config.json is the source of truth. render rebuilds CLAUDE.md, .claude/ and progress/ from it — idempotent, and without touching what you wrote.
navori.config.json
navori init per machine
The machine
Sessions that start outside a navori repo — a scratch dir, someone else's project, your home — have no harness at all. The global layer installs a floor of doctrine in ~/.claude, and steps aside on its own when the repo brings its own.
~/.claude · ~/.navori/global.json
navori global init per organization
The workspace
Org-wide defaults — quality gate, branch base, conventions — plus the Dominio: durable facts that span repos and fit in no single CLAUDE.md. A data model, a contract between services, a shared gotcha.
~/.navori/workspaces/<org>/
navori workspace · navori dominio Zero footprint without opt-in: without the matching init, navori has not written a single byte outside your repo.
How it works
Five cascading layers , one source of truth
Each layer composes on top of the previous one. You choose how much control you want at project level.
- navori
Core
The baseline navori owns: agents, skills, hooks and the doctrine blocks. Always present.
- preset
Preset
Stack-specific defaults — Next.js, NestJS, Astro, Expo, Medusa, monorepos and more. Missing yours? Scaffold a local one with `navori preset init`.
- org
Workspace
Org-wide rules shared across repos, plus the workspace Dominio.
- you
Project
What's unique to this repo in navori.config.json: quality gate, critical areas, legacy paths and the plugins you enable.
- engine
Adapters
Per-engine rendering for Claude Code, Codex, Cursor, Copilot and AGENTS.md — all on one pipeline.
navori.config.json
source of truth
{
"$schema": "https://ulisescm.github.io/navori-harness/schema/navori.config.v1.json",
"name": "your-app",
"engines": ["claude", "codex"],
"preset": "nextjs",
"branchBase": "main",
"commits": "conventional-es",
"qualityGate": {
"full": "pnpm lint && pnpm test"
},
"plugins": {
"engram": { "enabled": true },
"codegraph": { "enabled": true },
"tgrep": { "enabled": true },
"semgrep": { "enabled": true },
"gh": { "enabled": true }
},
"project": {
"criticalAreas": ["auth", "billing"],
"legacyPaths": ["src/legacy"]
},
"audit": { "mode": "always" }
} Live
See what each layer contributes
Click a layer to highlight what it contributes. Adapters materialize it in each engine's native format.
Pick a layer to see what it adds.
Engine adapters → .claude/ + CLAUDE.md
output
CLAUDE.md
- idioma-rol · formato-respuesta · tipado-fuerte · cierre-sesion core
- engram-protocol project
.claude/agents/
- orchestrator · implementer · reviewer · scout · … core
.claude/skills/
- verify-before-done · debug-failure core
- nextjs-app-router · nextjs-data-fetching preset
.claude/settings.json
- permisos + hooks base core
- hooks de plugins (engram, gh) project
.claude/hooks/quality-gate-pre-commit.sh
- qualityGate.fast heredado workspace
progress/
- current.md · history.md core
Toolbox
navori doesn't run the tools , it teaches the agent to use them
Every plugin is a bundle: its skill, its protocol block, its permissions and its doctor check. Enable the ones you want and the harness knows when to reach for them — and when not to.
acli external Jira
Reading and writing tickets through acli, from the right account. Ticket intake starts at the ticket, not at a description pasted by hand.
codegraph search Search by structure
A local AST graph of the repo. Where a symbol lives, who calls it and what breaks if you change it — one query instead of a round of greps and reads.
engram context Persistent memory
Decisions, root causes and conventions that survive session close and context compaction. You don't re-explain on Monday what you settled on Friday.
gh external GitHub
Issues, PRs and checks through gh. The agent reads the ticket, opens the PR in the repo's format, and knows how to read a red CI.
jscpd quality Duplication
Catches copy-paste before it becomes debt. An agent that can't see the code that already exists rewrites it; this is what shows it.
semgrep quality Security
Static analysis for dangerous patterns over the diff, wired into the reviewer's gate instead of sitting there as a step someone has to remember to run.
tgrep search Search by content
A trigram index over the repo. Which files hold this string, with the exit-code contract intact and no fight over which engine the machine happens to have.
navori generates the harness; it does not run grep, the tests or the linter on the agent's behalf. It dictates which tool to use and under what doctrine — the rest is executed by whoever owns it.
Observability
The harness that audits itself
Writing doctrine is easy; knowing whether anyone followed it isn't. navori audit answers the two questions nothing else does: where the tokens went, and which instructions nobody obeyed.
$ navori audit --session latest◇ navori-harness · 2026-09-12 ─────────╮│ 1 sesiones · 19 agentes ││ facturable 2.3M tok ││ arranque 346k tok ││ hallazgos 1 alto · 3 medio │╰──────────────────────────────────────╯→ report.md · report.json
Where the tokens went
Billable, startup, and per agent. The harness's cost stops being a hunch and becomes a number you can attack.
Which instructions nobody followed
Severity-ranked findings about real routing: the skill that got ignored, the search that took the expensive path, the delegation that never happened.
Three sources that don't replace each other
The event log the hooks write (what the harness did), the transcript (the only place tokens live), and the host's OTel events (which permission was approved, which skill was active).
Opt-in, one session at a time
With no prior --start there is no log to audit and navori observes nothing. The report lands as markdown and JSON under ~/.navori/audits/.
Commands
The whole lifecycle , command by command.
Every subcommand the CLI registers, grouped by when you actually need it. Each one has its own reference page.
Get started
- navori init Detects the stack, asks a few questions, and leaves the repo ready in a minute.
- navori adopt Takes a harness file you wrote by hand and puts it under management, without changing what it says.
- navori add Enables a plugin, or suggests which ones fit your stack.
- navori remove Disables a plugin and cleans up the blocks and scripts it left behind.
- navori preset Scaffolds a local preset when your stack has no official one, and wires it into the config.
- navori configure Edits the config without opening it by hand: quality gate, critical areas, legacy paths.
Render and update
- navori render Rebuilds every configured engine. Preview by default; --apply writes.
- navori sync Refreshes only the managed blocks. With --interactive you resolve conflicts block by block.
- navori update Detects that the bundle moved and tells you what would change first.
- navori migrations Applies config migrations between CLI versions.
- navori backup The safety net: every write leaves a restorable snapshot first.
Operate
- navori doctor Drift, broken plugins, invariants, conflicting foreign harness, and next steps.
- navori status At-a-glance snapshot: config, plugins, drift, and what to do next.
- navori scan Walks a monorepo and assigns a preset to each workspace.
- navori registry Which presets, plugins and engines this version ships.
- navori bench Times render over N runs and reports p50/p95.
- navori receipt
Scale
- navori global The per-machine base harness in ~/.claude, for sessions outside a navori repo.
- navori workspace Cross-repo config, and rendering the standard to every org repo.
- navori dominio The workspace knowledge base: the facts that span repos.
- navori ticket Tickets as versionable files inside the workspace.
Multi-engine
One config , five engines.
The same harness materialized in each one's native format, over a single render pipeline. Adding the sixth costs a declarative table, not a rewrite.
available
Claude Code
.claude/
agents · skills · hooks · settings.json with permissions
Codex
.codex/ + .agents/skills/
custom agents · skills · hooks · MCP servers
AGENTS.md
AGENTS.md
universal spec · read by Cursor, Codex, Gemini and Copilot
available
Cursor
.cursor/rules/
.mdc rules · project context
available
Copilot
.github/copilot-instructions.md
workspace instructions
Quickstart
From zero to productive in three steps
Install
No global install needed. Use npx.
npx navori init
Initialize
Answer a few prompts and get navori.config.json plus the rendered harness.
? Project › my-app ? Engine › claude ? Preset › nextjs ✓ Done — 5 created
Render whenever you change something
Edit the config, run render --apply, commit. Fully idempotent.
$ vim navori.config.json $ npx navori render --apply $ git add . && git commit
FAQ
What you're probably wondering
What if I already have a hand-rolled .claude/?
navori coexists. init detects your existing harness and only adds managed blocks marked with delimiters; your custom content stays put. If you want navori to take over one of your files, 'navori adopt' wraps it without changing a word of what it says. And if your setup already ships its own orchestration or SDD, blocks.exclude lets navori opt out of rendering those blocks so it never competes with yours.
Which engines does it support today?
Five, all shipped: Claude Code (.claude/), native Codex (agents, skills, hooks and MCP), universal AGENTS.md, Cursor (.cursor/rules/) and Copilot. They run on one render pipeline, so a fix reaches all of them at once instead of diverging in silence.
Do I need Claude Code to use navori?
No. navori is a scaffolder: it generates the harness and gets out of the way. The tool runs on any Node 20+ and renders to five different engines.
How invasive is it? Can I roll it back?
render previews by default: without --apply it never touches disk. When it does write, it leaves a restorable snapshot first via 'navori backup', and the write is atomic. Outside the repo nothing exists that you didn't ask for by name: no 'global init', no global layer; no '--start', no audit.
How do I upgrade my project when a new version ships?
npx navori@latest sync. It refreshes only managed blocks without touching your code; with --interactive you resolve block by block whatever you edited by hand. 'navori update' tells you what would change first.
Isn't running a multi-agent harness expensive?
Less than you'd think — and you no longer have to take that on faith: 'navori audit' tells you where a real session's tokens went. Each agent runs with its own model and effort — the muscle only where judgment lives, mechanical work on a light model — and every subagent returns its conclusion instead of a dump of what it read.