Netra Everywhere

Turn a project idea into a working, tested application with your coding agent.

Netra is a local harness for Codex, Claude Code, and OpenCode. Give it a plain-language idea or project files; it creates project-specific context and dynamic skills, guides the agent through planning and implementation, runs declared checks, and records the evidence locally.

Netra is not a hosted AI service. It has no Netra API key, account, telemetry, or remote project database. Your installed coding agent performs the reasoning and writes the code in your local workspace.

Why Netra

Most agent workflows start with generic instructions. Netra starts with the actual product:

• A user can provide only an idea; Build Mode creates an internal PROJECT_BRIEF.md rather than requiring a finished PRD.
• It forges only the skills the project needs: UI, API contracts, authentication, data schema, security review, and verification.
• It keeps implementation grounded in source-backed requirements, assumptions, phases, tasks, checks, and an append-only BRAIN.md activity log.
• It tells the active coding agent to build and verify the application. A skill package or a plan is never the final deliverable.
• It keeps risky actions explicit: deployment, production access, paid services, pushes, PRs, and destructive migrations require approval.

What Build Mode does

Your idea or files
       │
       ▼
Project brief + requirement clarification
       │
       ▼
Dynamic project-specific Netra skills
       │
       ▼
Source-backed plan, tasks, and verification checks
       │
       ▼
Coding agent builds a usable vertical slice, then remaining scope
       │
       ▼
Tests, bounded repairs, BRAIN evidence, and runnable code

Netra deliberately does not claim that an agent searched the web, queried a provider, ran a test, or deployed software unless that action actually happened and was recorded.

Install from GitHub

Requirements: Git, Node.js 24, and a supported coding-agent runtime.

From the root of the project you want to build:

npx --yes github:D0-6/Netra#main install . --target all
npx --yes github:D0-6/Netra#main doctor . --json

--target all installs:

.agents/skills/netra/         Portable Agent Skill for Codex and compatible runtimes
.claude/skills/buildmode/     Native Claude Code /buildmode command
.opencode/commands/buildmode.md  Native OpenCode /buildmode command

If npm cache permissions prevent npx from working, clone directly from GitHub:

git clone https://github.com/D0-6/Netra.git C:\Tools\Netra
node C:\Tools\Netra\bin\netra.mjs install . --target all

Build a project

Claude Code or OpenCode

/buildmode Build a polished web app where founders paste an idea, collect public evidence, compare similar products, and receive a cited novelty report.

Codex

Codex uses portable Agent Skills rather than a native slash-command registry:

Use $netra in build mode. Build a polished web app where founders paste an idea, collect public evidence, compare similar products, and receive a cited novelty report.

Build Mode should:

1. Create PROJECT_BRIEF.md when the user supplied only an idea.
2. Ask only material unanswered questions and record visible assumptions.
3. Forge dynamic skills supported by the project requirements.
4. Create a plan and verification checks.
5. Implement the project in the current workspace.
6. Run checks, make bounded repairs, and leave runnable code plus honest blockers.

Demonstrated comparison: static mockup vs. Netra build

Netra was evaluated by asking two coding-agent workspaces to make the same ResearchRadar concept: a tool where founders submit an idea, gather public evidence, compare similar products, and receive a novelty report.

Evidence	Baseline build (without Netra)	Netra Build Mode build
Project shape	Three static browser files	Node server, browser client, SQL schema, project documents, and tests
UI	Stronger visual dashboard and motion	Simpler but live, data-driven screens
Accounts and protected data	Not implemented	Sign-up, sign-in, sign-out, session-gated dashboard, workspace ownership
Product records	Hard-coded demo cards	Create workspaces, projects, analysis runs, and Markdown reports through APIs
Evidence handling	Fabricated sample evidence in the interface	Explicitly says no evidence exists when no provider returned it
Automated checks	None	npm test and npm run check passed during the evaluation
Public-web search	Not implemented	Provider interface exists, but no live provider was configured

The result is deliberately not presented as a finished production product. The Netra build is the more credible engineering prototype because it has a real request flow and does not invent research results. Its most important remaining gap is also clear: without a configured server-side research provider, it cannot yet search GitHub, Product Hunt, papers, or the wider web.

The full reproducible audit, scope, commands, and known limitations are in the ResearchRadar comparison example.

Project files and local state

Netra works with an idea alone, but it gains precision from files such as README.md, PRD.md, FEATURES.md, SYSTEM_ARCHITECTURE.md, and design notes.

PROJECT_BRIEF.md              Created from an idea when needed
BRAIN.md                      Append-only plans, decisions, changes, and validation evidence
.netra/
  PROJECT_MODEL.md            Deterministic document index
  REQUIREMENTS.md             Questions, answers, assumptions, and approval state
  CAPABILITY_MAP.md           Product capabilities and boundaries
  MODULE_SELECTION.md         Why a dynamic skill was selected
  verification.json           Explicit local commands allowed for verification
  phases/ and tasks/          Durable implementation plan and task state
  evidence/                   Recorded verification evidence
.agents/skills/netra-*/       Generated project-specific skills

Main commands

Command	Purpose
netra doctor .	Inspect installation, runtime availability, and migration state without changing files.
netra init .	Initialize durable local Netra state.
netra ingest <files> --project .	Index the actual project requirements.
netra interview .	Create source-bound questions for material unknowns.
netra forge . --target both	Create the project-specific skill packages.
netra plan, phase-define, task-add	Create source-backed phases and bounded tasks.
netra execute <phase>	Create an implementation handoff for the active coding agent.
netra verify .	Run only declared local verification commands.
netra runtime-start <phase> --approve-runtime	Explicitly launch the local Codex CLI adapter.
netra wave-*	Plan, launch, review, and merge bounded isolated task workers.
netra review, ship-check, commit	Prepare evidence-led review and an approval-gated local commit.

See HARNESS_SPEC.md for the complete V1 command contract and ROADMAP.md for work that is not implemented.

Safety model

• Netra never stores user API keys or sends project content to a Netra service.
• verify accepts only explicit local command arrays from .netra/verification.json.
• Generated skills never overwrite unmarked user-owned skills.
• Requirements approval becomes stale when source files change.
• Commits, worker launches, worktrees, state merges, runtime launches, and security scans require explicit flags.
• Netra does not push, open pull requests, deploy, run production migrations, or access production systems by default.

Built with Codex and GPT-5.6

Netra was designed and implemented with Codex powered by GPT-5.6. Codex was used to shape the agent-harness architecture, implement dynamic skill forging and Build Mode, add state and safety gates, build the verification workflow, and create the automated test suite.

The project remains local-first: Netra supplies the workflow and project context; the coding agent installed by the user performs the actual implementation in the user’s environment.

Development

npm install
npm test
npm run preflight

Run the local CLI during development:

node .\bin\netra.mjs install C:\Projects\my-product --target all

Read CONTRIBUTING.md before changing the public command contract. A command enters V1 only with a passing test in the same commit.

License

MIT © 2026 D0-6. See LICENSE.