Nova AI — what it is, why it exists

The PM tool that ships the work

Describe a ticket. AI reads your codebase, writes the fix, opens a merge request. You approve. Done.

Features designed

—

Specs ready to build

Build phases

~$40

Infra / month

The pitch

In three paragraphs

Nova AI is a project management tool with a self-executing ticket loop. You describe what needs to happen in plain English. The AI reads the codebase, writes the fix, opens a merge request, and waits. You review. You approve. It merges and ships.

It replaces Monday.com and Jira for technical teams who want a tool that actually moves work forward — not a glorified spreadsheet that someone has to nag people into updating.

Nova AI is autopoietic — it builds itself. New features are logged as tickets in Nova AI, and Nova AI implements them. Every shipped feature is a working proof of the product.

Autopoietic

The tool builds itself using the same loop it manages. Every new feature shipped by Nova AI is a proof of the product.

Positioning

How it's different

vs Monday.com

Monday is a dashboard for humans to coordinate. Nova AI is an agent that executes tickets and opens merge requests. No ceremony, no drag-and-drop.

vs Copilot

Copilot assists a developer writing code. Nova AI takes a ticket from "todo" to "merge request ready" — reading the codebase, choosing the approach, writing the fix, committing it.

Foundation

Six core principles

Human approval gate is non-negotiable. The AI never merges to main without explicit human sign-off. MR must be approved by a human in GitLab. This cannot be bypassed, configured away, or overridden by the AI.

AI permissions are structural, not policy-based. Safety comes from scoped API tokens and absent tools — not from trusting the AI to behave. The agent literally cannot call merge_mr, delete_file, or any admin function because those tools don't exist.

Nothing stored in Nova AI that belongs elsewhere. Code in GitLab. Docs in Google Drive + GitLab docs repo. Snippets in the snippets repo. To-dos in the todos project. Nova AI's Firestore holds only execution state, notifications, planning threads, preferences, and audit log — never raw content.

Pay for what you use. No per-seat pricing. API costs are usage-based. ~$40/month infrastructure on the team's own GCP. Scale to zero when idle.

Everything on GCP. One billing account, one console, no third-party hosting dependencies. Cloud Run, Firestore, Vertex AI, Cloud Tasks, Cloud Build, Memorystore, Secret Manager.

Best AI for each task. Claude Sonnet 4.6 for code execution and reasoning. Claude Haiku for classification and cheap tasks. Gemini 2.5 Flash via Vertex AI for document Q&A and stale detection. No one-size-fits-all model.

Mechanics

How the loop works

Describe the ticket in plain English. The AI helps refine and asks clarifying questions if needed.

AI reads the context — codebase, related Drive docs, similar past tickets.

AI writes the fix on a feature branch and opens a merge request.

You review and approve. GitLab is the source of truth — same review flow you already use.

It ships. Cloud Build picks up the merge and deploys.

The AI literally cannot merge without you. The merge_mr tool doesn't exist in its toolkit. Safety is structural, not a policy that can be overridden.

Delivery roadmap

The build phases

Phase 1Workspace

Team Workspace

Issues list/board, cycles, planning threads, MR review reminders, notification inbox, auth, admin. The daily workspace that replaces Monday.com.

Phase 2Execution

AI Execution

AI ticket execution, human approval gate, execution persistence, AI revision loop, AI-assisted ticket creation. Proves the core concept.

Phase 3Intelligence

Intelligence & Data

Roadmap, dashboard, Gemini doc Q&A, stale detection, Miro, engine release notifications, QA briefings, Game Data Hub.

Phase 4Operations

Operations

Absence management, reviewer pool enhancements, QA agent, sound management. Team operations automation.

Phase 5Platform

Extensible Platform

Self-deploy pipeline, extensible core, ⌘K palette, slash commands, multi-agent architecture. Nova AI ships itself from this phase forward.

Under the hood

Tech stack at a glance

Frontend

Next.js + TypeScript, Tailwind, shadcn, Tiptap (free core).

Backend

Fastify on Cloud Run.

AI

Claude Sonnet 4.6 (code), Claude Haiku (classification), Gemini 2.5 Flash (docs).

Data

Firestore (metadata only — code lives in GitLab, docs in Drive).

Infra

GCP — Cloud Run, Cloud Tasks, Memorystore, Secret Manager, Cloud Build.

Hosting

Single GCP project. One console, one bill, no third-party hosting dependencies.

Runs on the team's own GCP. ~$40/month infrastructure. No per-seat fees.

Where to go from here

See it in action

10-chapter product tour

→ walkthrough

Browse the specs

Full ticket specifications