A product studio for hackathon teams.

Research the possibilities. Make the decisions. Leave with a plan your team can build.

Try locally · The workflow · Developer guides

5 ranked directions	8 explained scores	10 Markdown files	5 AI providers
Compare before committing	See reasoning and trade-offs	PRD, AGENTS, and a build handoff	Bring your own key

From brief to build plan

Real application captures use explicitly labeled offline fixtures. They demonstrate the workflow, not live model or research quality.

flowchart LR
    A[Challenge + team + time] --> B[Generate ideas]
    B --> C[Research + critique + revise]
    C --> D[Rank top five]
    D --> E{Your decision}
    E -->|Refine| C
    E -->|Select| F[Review stack]
    F -->|Edit| F
    F -->|Confirm| G[Markdown handoff]
    G --> H[Project assistant]

Loading

Explore with evidence	Keep creative control	Build from a shared plan
Competitors, source links, critiques, risks, effort estimates	Refine ideas with a prompt; edit and approve the stack	Human-readable requirements and dedicated agent instructions
Team skills, deadline, budget, and judging criteria shape the result	Decisions stay with the team at each approval gate	Persistent assistant uses the plan, research, team, and chat

Important

Scores are advisory estimates. Search cannot prove uniqueness or guarantee a winning entry. Evidence and reasoning stay visible for your review.

Your next commit starts here

Product definition	Technical decisions	Build execution
PRD.md	TECH_STACK.md	AGENTS.md
RESEARCH.md	ARCHITECTURE.md	IMPLEMENTATION_PLAN.md
README.md	DATABASE.md · API.md	SETUP.md

Preview · Markdown source · Copy · Individual downloads · Reproducible ZIP

Tip

Documents come from one approved snapshot. Reading never regenerates them. Upstream changes invalidate affected approvals and documents; earlier snapshots remain in History.

Try it locally

Clone the repository, then start the explicit offline evaluation with Docker Compose v2:

git clone https://github.com/Muneer320/exHacker.git
cd exHacker
docker compose -p exhacker-verify -f compose.verify.yml up --build --wait

Open localhost:3200. No .env or API keys needed.

Runs for real	Uses offline fixtures
Frontend, API, PostgreSQL, migrations, jobs, persistence, exports	Idea content, scores, documents, demonstration assistant

A five-minute walkthrough

1. Enter a challenge, deadline, and teammates with their skills.
2. Compare five directions and refine one with your own prompt.
3. Select it, edit a technology choice, and confirm the stack.
4. Preview PRD.md and AGENTS.md; copy or download the handoff.
5. Ask the assistant how to split the work. Reload to see saved history.

Natural-language stack revisions need live AI; manual edits work offline. Stop while retaining evaluation data:

docker compose -p exhacker-verify -f compose.verify.yml down

Connect live AI and web research

Copy the root .env.example to .env: cp .env.example .env on macOS/Linux or Copy-Item .env.example .env in PowerShell. Configure your Tavily key for search:

AI_ACCESS_MODE=byok
AI_LEASE_SECONDS=1800
TAVILY_API_KEY=your-search-api-key
STUDIO_DEMO_MODE=false

docker compose up --build --wait

Open localhost:3000, click Connect AI, and enter your provider key in the dedicated form. This development stack persists SQLite in its own volume. The separate app on port 3200 always uses verification fixtures.

Use the API-key testing walkthrough to check the full workflow, switching, expiry, and recovery. Missing credentials produce visible errors and preserve saved work.

Develop without Docker

Use Python 3.12 (minimum 3.11), uv, and Node.js 22+. Configure the root .env, then start two terminals:

# Terminal 1
cd backend
uv sync --frozen --extra dev
uv run uvicorn app.main:app --reload

# Terminal 2
cd frontend
npm ci
npm run dev

The frontend proxy defaults to http://127.0.0.1:8000. Override server-side BACKEND_URL in the frontend process or frontend/.env.local. The root .env is not automatically a Next.js environment file. See configuration.

Your choice of intelligence

OpenAI	Anthropic	Google	xAI	Groq
GPT-6 Luna	Claude Haiku 5.5	Gemini 3.8 Flash	Grok 4.7	GPT OSS 120B / 20B

Choose a model → Connect a temporary key → Switch future operations → Keep your artifacts

See the temporary AI session controls

Real interface; no entered API key. This capture does not establish paid-provider access.

Note

AI keys live in backend RAM under a default 30-minute lease. Expiry, removal, or restart pauses affected jobs for reentry. Reloading clears this tab's lease references; saved projects remain. Tavily stays operator-managed. Account quotas can be smaller than model context limits. BYOK design and limitations.

Under the workspace

flowchart LR
    UI[Next.js + React] --> Proxy[Same-origin proxy]
    Proxy --> API[FastAPI + embedded worker]
    API <--> DB[(Projects / jobs / artifacts / chat)]
    API --> AI[Request-scoped HTTPX / five providers]
    API --> Web[Tavily research]
    API --> MD[Validated specification / Markdown compiler]
    MD --> DB

Loading

Interface	Application	Recovery
Next.js 16.4 · React 19 · TypeScript	FastAPI · Pydantic · SQLAlchemy · Alembic	Durable jobs · revision checks · approval gates
Responsive black, blue, and lime workspace	SQLite locally · PostgreSQL verified in Docker	Atomic publication · leases · saved research checkpoints

What survives a failure?

Failed operations retain the previous successful context. Stale tabs cannot overwrite newer revisions. Saved candidate/research checkpoints let unchanged idea jobs resume within their reuse window. AI leases never enter the database; restarts require credential reentry. Recovery can repeat paid work; external billing is not exactly once. See the architecture.

The developer desk

Run and contribute	Understand and operate
Configuration and environment variables	Architecture and directory map
Tests, browser checks, and live-key walkthrough	BYOK sessions and recovery
Contributing	Deployment and operations
MIT license	Security and privacy

Warning

Active development; public-hosting validation remains open. BYOK needs one persistent API process with its embedded worker. Review HTTPS, PostgreSQL, backups, abuse controls, live-provider quality, dependency advisories, and load testing before hosting. Start with the deployment guide.

Caution

Save your workspace key privately. Anyone holding it can access the workspace; losing it loses access. Project text and chat persist and may be sent to selected providers. Enter credentials only in the dedicated AI session form. Security boundaries.

Created by Muneer Alam · MIT license · Contributions welcome