🚀 Hack Day Starter — Run Local AI Models on Any Laptop or PC

Hardware-aware local model recommender and zero-dependency TypeScript starters for Ollama.
Built for the Hacktoberfest 2026 "Build for a Friend" Weekend Challenge, Hack Day Starter profiles your hardware (RAM, GPU, CPU, SSD headroom), matches your machine with a verified open-weight AI model fitted to your system memory and compute limits, and generates an immediate, production-ready TypeScript project with zero runtime dependencies.

📚 Technical Documentation & Guides

📖 Verified Model Registry: All 23 tracked entries with exact manifest byte sizes, context token windows, lifecycle states, and verified capabilities.
⚡ Ollama Hardware Requirements Guide: In-depth engineering analysis of Apple Silicon Metal, NVIDIA CUDA VRAM, CPU bottlenecks, and KV cache calculations.
🛠️ Building Local Tool-Calling Agents: Native Ollama tool calling, multi-step agent loops, AST arithmetic security, and verified models.
⚙️ Deterministic Sizing Logic: Mathematical scoring breakdown, binary safety gates, and 1.5 GB buffer policy.

⚡ Quick Answer: How to Run a Local AI Model on Your Device

If you are asking "How do I run an open-source AI model locally on my computer?", here is the 3-step process:

Install Ollama (the local AI runtime):

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh
# Windows: Download installer from https://ollama.com/download

Match the right model to your RAM:
- 8 GB RAM: Run phi4-mini (2.5 GB) or qwen3.5:4b (3.4 GB)
- 16 GB RAM: Run gemma4:e4b (6.6 GB) or qwen3.5:9b (6.6 GB) or gemma4:12b (7.7 GB)
- 24 GB – 32 GB RAM: Run gemma4:12b (with 256K context) or qwen3.8:27b (18 GB)
- 64 GB+ RAM: Run qwen3-coder-next (52 GB) or large MoE models
Pull and execute:
```
ollama run gemma4:12b
```

The catch? Wiring up streaming, handling conversational state, managing function/tool calls, avoiding disk swap death, and writing boilerplate code takes hours.
Hack Day Starter automates this end-to-end.

🎯 The "Build for a Friend" Challenge (Hacktoberfest 2026)

Hack Day Starter was built specifically for the Hacktoberfest 2026 "Build for a Friend" Weekend Challenge. During hackathons, weekend sprints, and local AI prototyping, developers lose the first 2–4 hours fighting common pitfalls:

💥 Out-of-Memory (OOM) Crashes: Downloading a 14 GB or 20 GB model on an 8 GB or 16 GB laptop triggers aggressive swap thrashing, freezing the machine or crawling at 0.2 tokens/second.
📦 Stale Model Catalogs: Outdated tutorials recommend retired 2023-era models (like Llama 2 7B) instead of modern state-of-the-art architectures (Gemma 4, Qwen 3.5, Phi-4) that offer 256K context and native tool calling.
💾 Disk Space Blindspots: Pulling models without checking SSD headroom can stall multi-gigabyte downloads halfway through.
🔌 Boilerplate & Runtime Frustration: Connecting to local REST daemons, handling chunked NDJSON streams, configuring tool execution loops, and debugging raw ECONNREFUSED connection crashes drain developer momentum.

Hack Day Starter was created to solve this for a friend:

Detect your hardware specs (RAM, GPU acceleration, OS, free storage).
Deterministically recommend verified models with realistic memory requirements.
Choose your starter archetype (Local Chat or Tool-Calling Agent).
Download a clean, ready-to-run TypeScript project ZIP and start building in under 60 seconds.

📊 Hardware Sizing Matrix: What Model Can Your Computer Run?

Memory calculations must account for the model weights, the KV cache (which scales with context length), and operating system overhead (~2.5–4 GB):

Your Available RAM	Recommended Models	Ollama Tag	Model Size	Context Window	Native Tool Calling	Experience / Speed
8 GB RAM (Standard Laptop)	Phi-4 Mini Nemotron 3 Nano Qwen 3.5 4B	`phi4-mini` `nemotron-3-nano:4b` `qwen3.5:4b`	2.5 GB 2.8 GB 3.4 GB	128K 256K 256K	✅ ✅ ✅	⚡ Fast (30–60 tok/s on M-series or modern CPU)
16 GB RAM (Dev Laptop)	Gemma 4 E4B Qwen 3.5 9B Gemma 4 12B	`gemma4:e4b` `qwen3.5:9b` `gemma4:12b`	6.6 GB 6.6 GB 7.7 GB	128K 256K 256K	✅ ✅ ✅	🚀 Optimal balance of intelligence, speed, and context
24 GB – 32 GB RAM (Pro Mac / RTX PC)	Gemma 4 12B (Full KV) Qwen 3.8 27B GPT-OSS 20B	`gemma4:12b` `qwen3.8:27b` `gpt-oss:20b`	7.7 GB 18 GB 14 GB	256K 256K 128K	✅ ✅ ✅	🧠 High reasoning power, complex coding & agent tasks
64 GB+ Unified / VRAM (Workstation)	Qwen 3 Coder Next Qwen 3.6 35B	`qwen3-coder-next` `qwen3.6:35b`	52 GB 23 GB	256K 256K	✅ ✅	🏆 Full frontier-grade local intelligence

Note

Unified Memory (Apple Silicon M1–M6): macOS dynamically shares RAM between CPU and GPU. A 16 GB Mac can dedicate ~12 GB directly to model weights and KV cache, making it one of the most efficient local AI platforms available.

🏗️ System Architecture

               ┌───────────────────────────────────────────────┐
               │         Developer Laptop Specifications       │
               │  • RAM (GB)      • GPU (Apple / NVIDIA / CPU) │
               │  • Operating OS  • Free Disk Space Headroom   │
               └───────────────────────┬───────────────────────┘
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │             Hack Day Starter (App)            │
               │  • Deterministic hardware compatibility score │
               │  • Verified Model Registry (Oct 2026 data)    │
               │  • Authoritative vs Estimated memory bounds   │
               │  • Free disk space safety gate                │
               └───────────────────────┬───────────────────────┘
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │           Verified Model Selection            │
               │  • Exact Ollama tag (e.g. gemma4:12b)         │
               │  • Verified native capabilities (tools/vision)│
               │  • Starter type: Local Chat vs Agent          │
               └───────────────────────┬───────────────────────┘
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │      Generated Starter Project (ZIP)          │
               │  • Zero runtime dependencies (native fetch)   │
               │  • Direct /api/chat streaming client          │
               │  • Built-in Ollama & Model Readiness check    │
               │  • Safe recursive-descent math tool           │
               │  • Deterministic in-memory generation         │
               └───────────────────────┬───────────────────────┘
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │          Local Ollama Runtime Daemon          │
               │             http://localhost:11434            │
               └───────────────────────┬───────────────────────┘
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │          Open-Weight Model Weights            │
               │         gemma4:12b / qwen3.5:9b / ...         │
               └───────────────────────┬───────────────────────┘
                                       │
                         ┌─────────────┴─────────────┐
                         ▼                           ▼
                 Direct Response            Tool Call Execution
                                            (src/tools/calculator.ts)
                                                     │
                                                     ▼
                                            Synthesized Result

Component Roles & Boundaries

Component	Role	What It Does / Does Not Do
Ollama	Local Daemon	Runs locally on port `11434`. Pulls weights, manages GGUF quants, and executes hardware-accelerated tensor math on Metal/CUDA/Vulkan.
Open-Weight Models	Intelligence Layer	Neural network weights (e.g., `gemma4:12b`, `qwen3.5:9b`). Runs completely offline inside Ollama.
Hack Day Starter	Generator & Recommender	Profiles hardware, filters models, and builds custom starter ZIPs in-memory. Does not proxy inference or transmit prompts.
Generated Project	Your Application	Independent TypeScript codebase. Runs directly on your machine with zero external SDKs or telemetry.

🛠️ Starter Archetypes

1. Local Chat (`hack-day-starter-chat-<model>`)

Best for: Chatbots, personal offline copilots, document Q&A, and conversational prototyping.
Key Capabilities:
- Zero runtime dependencies: Uses native Node.js 18+ global fetch.
- Interactive CLI loop: Multi-turn conversational memory with seamless terminal I/O.
- Smart readiness check: Verifies the Ollama daemon and local model existence before prompting the user.
- Graceful shutdown: Clean EOF handling on Ctrl+C or typing exit.

2. Autonomous Tool-Calling Agent (`hack-day-starter-agent-<model>`)

Best for: Multi-step autonomous agents, calculation engines, and structured function execution.
Strict Safety Gate: Only selectable when the model is verified to support native tool calling (model.capabilities.tools === true).
Key Capabilities:
- Autonomous ReAct Loop: Model generates structured tool_calls → runtime executes tool → sends { role: "tool", content: result } back to model → model synthesizes natural language conclusion.
- Safe Math Engine: Ships with calculator.ts, a pure recursive-descent arithmetic parser (0% eval(), 0% new Function()).
- Extensible Architecture: Add new tools in minutes using standard JSON Schema definitions.

💻 Real End-to-End Execution Trace: Gemma 4 12B Agent

Here is the exact terminal output from the generated tool-calling agent running against local Ollama with gemma4:12b:

$ npm run dev

=================================================
🤖 Hack Day Starter — Tool-Calling Agent
🎯 Model: gemma4:12b (Native Tool-Calling Verified)
🛠️  Active Tools: [calculate]
📡 Ollama Host: http://localhost:11434
=================================================

✨ Agent ready! Try asking complex calculations or math problems:
Example: 'What is 342 multiplied by 19, and then divide by 4?'
Type 'exit' or press Ctrl+C to quit.

User: What is 342 multiplied by 19, and then divide by 4?
Agent: Thinking...
⚙️  [Tool Call] calculate({"expression":"(342 * 19) / 4"})
📋 [Tool Result] 1624.5
Agent: Thinking...
Agent: 342 multiplied by 19, divided by 4, equals 1624.5.

What Happened Behind the Scenes:

User Prompt: Submitted a natural-language multi-step arithmetic request.
Autonomous Tool Call: gemma4:12b recognized it required a calculation and emitted a structured tool call: calculate({"expression": "(342 * 19) / 4"}).
Safe Evaluation: src/tools/calculator.ts evaluated the expression mathematically without invoking any shell or arbitrary JavaScript execution.
Tool Context Feeding: The output 1624.5 was sent back to the model as a tool role message.
Final Synthesis: Gemma 4 synthesized the final, human-readable answer.

🔍 Verified Model Registry (Live Ollama Source Snapshot)

Unlike static blog posts that become outdated within weeks, Hack Day Starter uses a Verified Model Registry tracking 23 model entries categorized by capability verification, runtime testing status, and lifecycle. The complete catalog with manifest byte sizes and interactive filters is browsable on the Verified Model Registry page:

Model	Creator	Ollama Tag	Display Size	Context	Tools	Verification Status
Gemma 4 12B	Google	`gemma4:12b`	7.7 GB	256K	✅	Official Spec: 16 GB
Gemma 4 E4B	Google	`gemma4:e4b`	6.6 GB	128K	✅	Official Spec: 8 GB
Llama 3.2 3B	Meta	`llama3.2:3b`	2.0 GB	128K	✅	Runtime Verified
Llama 3.1 8B	Meta	`llama3.1:8b`	4.9 GB	128K	✅	Runtime Verified
Qwen 3.5 9B	Alibaba	`qwen3.5:9b`	6.6 GB	256K	✅	Source Verified
Qwen 3.5 4B	Alibaba	`qwen3.5:4b`	3.4 GB	256K	✅	Source Verified
Qwen 2.5 Coder 7B	Alibaba	`qwen2.5-coder:7b`	4.7 GB	32K	✅	Source Verified
Mistral NeMo 12B	Mistral	`mistral-nemo:12b`	7.1 GB	128K	✅	Source Verified
Qwen 3.8 27B	Alibaba	`qwen3.8:27b`	18 GB	256K	✅	Source Verified
Qwen 3.6 35B	Alibaba	`qwen3.6:35b`	23 GB	256K	✅	Source Verified
GPT-OSS 20B	Community	`gpt-oss:20b`	14 GB	128K	✅	Source Verified
Phi-4 Mini	Microsoft	`phi4-mini`	2.5 GB	128K	✅	Source Verified
Nemotron 3 Nano 4B	NVIDIA	`nemotron-3-nano:4b`	2.8 GB	256K	✅	Source Verified
LFM 2.5 8B	Liquid	`lfm2.5:8b`	5.2 GB	32K	❌	Source Verified
Devstral Small 2	Mistral	`devstral-small-2:latest`	15 GB	32K	✅	Source Verified
DeepSeek R1 1.5B	DeepSeek	`deepseek-r1:1.5b`	1.1 GB	128K	❌	Source Verified

🚦 Ollama & Model Readiness UX (Zero Stack Traces)

First-time local AI developers frequently run into confusing crashes:

FetchError: connect ECONNREFUSED 127.0.0.1:11434
model 'xyz' not found, try pulling it first

Generated projects ship with an integrated Readiness Check (src/readiness.ts). Before any prompt is processed, it checks:

Is Ollama Running?
If the daemon is offline, it prints a clean, formatted guide:

===============================================================
⚠️  OLLAMA IS NOT REACHABLE
---------------------------------------------------------------
Could not connect to Ollama at: http://localhost:11434

To run this starter, please:
1. Install Ollama: https://ollama.com/download
2. Start the Ollama application or run 'ollama serve'
3. Restart this starter: npm run dev
===============================================================

Is the Model Installed?
If the weights are missing, it provides the exact copy-paste command:

===============================================================
⚠️  MODEL NOT INSTALLED: gemma4:12b
---------------------------------------------------------------
Ollama is running, but 'gemma4:12b' is not downloaded yet.

To download this model, run:
  ollama pull gemma4:12b

Once download finishes, restart this starter:
  npm run dev
===============================================================

🔒 Security & Privacy Commitments

No External API Keys: Inference executes locally through your local Ollama instance. Prompts, code, and documents are processed on your machine rather than sent to third-party cloud APIs.
Deterministic & Safe Generation: The generator uses pure template composition. It does not invoke opaque third-party cloud LLMs to write code.
No Dynamic Code Evaluation: src/tools/calculator.ts strictly rejects any input containing letters, semicolons, backticks, or non-arithmetic characters. It never touches eval() or new Function().
Registry Enforcement: The API only generates starters for models registered in VERIFIED_MODEL_REGISTRY. Arbitrary, unverified model tags are rejected.

🚀 Step-by-Step Installation & Quickstart

Method 1: Using the Web App Profiler (Recommended)

# 1. Clone this repository
git clone https://github.com/TechGenDM/hack-day-starter.git
cd hack-day-starter

# 2. Install dependencies
npm install

# 3. Launch the web profiler
npm run dev

# 4. Open http://localhost:3000 in your browser
#    - Review your hardware recommendations
#    - Select Chat or Agent starter
#    - Click "Download Starter (.ZIP)"

Method 2: Running Your Downloaded Starter Project

# 1. Unzip your generated project
unzip hack-day-starter-agent-gemma4.zip -d my-ai-agent
cd my-ai-agent

# 2. Pull the model weights locally
ollama pull gemma4:12b

# 3. Install dev dependencies (TypeScript & tsx)
npm install

# 4. Start your local AI!
npm run dev

🧪 Comprehensive Automated Test Suite (36/36 Passing)

The codebase includes an extensive automated test suite covering recommendation logic, live data syncing, generator determinism, AST security auditing, runtime readiness diagnostics, score transparency, and starter setup flow contracts:

npm test

✔ R1: Both starter types ship the readiness module and use it in index.ts
✔ R2 (chat): Ollama unavailable → friendly message, download URL, no raw error
✔ R2 (agent): Ollama unavailable → friendly message, download URL, no raw error
✔ R3 (chat): model missing → exact 'ollama pull qwen3.5:9b', no substitution
✔ R3 (agent): model missing → exact 'ollama pull gemma4:12b', no substitution
✔ R4: Chat — ready → readiness message, then unchanged chat behavior with exact tag
✔ R5: Agent — ready → unchanged model → calculator tool → result → model loop
✔ R6: readiness module matches exact tags only (':latest' normalization for bare names)
✔ A: Gemma 4 E4B metadata matches verified source snapshot
✔ B: Gemma 4 12B metadata matches verified source snapshot
✔ C: Qwen 3.5 4B metadata matches verified source snapshot
✔ D: Qwen 3.5 9B metadata matches verified source snapshot
✔ E: GPT-OSS 20B metadata matches verified source snapshot
✔ F: Capability fields match source snapshot
✔ G: Ollama display sizes and short digests match source snapshot
✔ H: Current model discovery finds newly added model families including Qwen 3.8
✔ I: Retired / unavailable models are never recommended
✔ J: Stale metadata cannot be presented as 'current' or recommended
✔ K: 24 GB Apple Silicon recommends only models that pass current hardware policy
✔ L: 8 GB CPU-only does not receive artificially inflated model-capacity bonuses
✔ M: Tool-calling selection depends on verified capability data
✔ N: Discrepancy detector flags discrepancies when curated registry disagrees with observations
✔ O: Runtime verified models and runtime pending models are strictly distinguished
✔ P: Recommendation transparency and score factor accuracy
✔ A: Chat starter generation produces standard file set
✔ B: Agent starter generation produces tool-calling file set
✔ C: Exact selected Ollama tag is injected into config, code, and README
✔ D: README contains exact setup commands in sequence
✔ E: Agent generation rejects a no-tools model with exact error message
✔ F: Generated file paths are valid relative paths and JSON files parse cleanly
✔ G: Generated project is 100% deterministic across multiple runs
✔ H: Retired, unverified, or arbitrary models cannot bypass safety gates
✔ I: Generated templates do not access invalid readline.clearLine or cursorTo
✔ J: Calculator tool security audit — zero eval, zero Function, strict validation, correct arithmetic
✔ K: Starter API route enforces strict model safety and capability verification
✔ L: Starter Workbench setup flow prerequisites and command contract

ℹ tests 36
ℹ suites 0
ℹ pass 36
ℹ fail 0

❓ Frequently Asked Questions (FAQ)

Q1: Can I run an AI model locally without a dedicated GPU?

Yes. Modern open-weight small language models (such as phi4-mini at 2.5 GB or qwen3.5:4b at 3.4 GB) run efficiently on standard modern x86/ARM CPUs using Ollama's SIMD and AVX-512 optimizations. On Apple Silicon (M1–M6), the CPU and GPU share high-bandwidth Unified Memory, giving desktop-grade performance on thin laptops.

Q2: How much RAM do I need for local AI?

Minimum: 8 GB RAM allows comfortable execution of 2B–4B parameter models.
Sweet Spot: 16 GB RAM runs 7B–12B models (e.g., gemma4:12b or qwen3.5:9b) at full speed with generous context windows.
Power User: 32 GB–64 GB RAM unlocks 27B–35B parameter reasoning models and coding specialists.

Q3: Why run models locally instead of using cloud APIs like OpenAI or Anthropic?

Local Privacy & Data Control: Prompts, code, and documents run directly on your machine instead of being transmitted to third-party cloud servers.
Zero Incurred Usage Fees: No subscription tiers, no per-token billing, and no credit card required.
Offline Resilience: Functions completely disconnected from external networks (airplanes, field work, secure offline environments).
No Cloud Rate Limits: Requests are bound only by your local hardware capabilities, without artificial vendor request caps or token-per-minute throttling.

Q4: What is the difference between Ollama and Hack Day Starter?

Ollama is the runtime engine (analogous to the Node.js runtime or Docker daemon) that manages model execution.
Hack Day Starter is the developer bootstrapper that profiles your machine, recommends the best model for your exact hardware, and scaffolds a complete TypeScript application with zero runtime dependencies.

Q5: How does tool-calling work without an internet connection?

The open-weight model is trained to recognize when a query requires a tool. It outputs a structured JSON object specifying the tool name and arguments. Your local TypeScript runtime executes the function on your machine (e.g., querying a local database or running a math parser) and returns the output to the model to generate the final response.

📁 Repository Structure

├── app/
│   ├── api/recommend/route.ts      # Hardware scoring & model recommendation API
│   ├── api/starter/route.ts        # Starter generation & model safety gating API
│   ├── globals.css                 # Dark theme & UI layout styling
│   ├── layout.tsx                  # Root layout & SEO metadata
│   └── page.tsx                    # Hardware profiler, model picker & ZIP downloader
├── lib/
│   ├── types.ts                    # Core TypeScript interfaces & registry schemas
│   ├── registry.ts                 # Curated Verified Model Registry
│   ├── memory-calculator.ts        # Authoritative vs estimated memory heuristics
│   ├── recommend.ts                # Deterministic recommendation scoring engine
│   ├── sources/
│   │   └── ollama.ts               # Ollama library observation scraper & parser
│   └── starter/
│       ├── types.ts                # Starter project types
│       ├── generate.ts             # In-memory deterministic project generator
│       └── templates/
│           ├── chat.ts             # Zero-dependency Local Chat template
│           ├── agent.ts            # Autonomous Tool-Calling Agent template
│           └── readiness.ts        # Ollama daemon & model readiness check template
├── scripts/
│   └── sync-model-registry.ts      # Automated registry discovery & discrepancy checker
├── tests/
│   ├── recommend.test.ts           # Tests A–P: 16 recommendation, transparency & registry tests
│   ├── starter.test.ts             # Tests A–L: 12 starter generation, security & setup flow tests
│   └── readiness.test.ts           # Tests R1–R6: 8 readiness & runtime integration tests
├── package.json
├── LICENSE                         # MIT License
└── README.md

🤝 Contributing

Contributions to improve hardware profiling heuristics, expand verified model registry snapshots, or add new starter templates are welcome!

Fork the repository
Create a feature branch: git checkout -b feature/model-sync
Ensure all tests pass: npm test
Submit a Pull Request

📄 License

Distributed under the MIT License. See LICENSE for details.
Built with ❤️ for Hacktoberfest 2026 — Weekend Challenge: Build for a Friend.

Name		Name	Last commit message	Last commit date
.agents		.agents
app		app
lib		lib
public		public
scripts		scripts
tests		tests
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json

Navigation Menu

Repository navigation

About

Repository files navigation

🚀 Hack Day Starter — Run Local AI Models on Any Laptop or PC

📚 Technical Documentation & Guides

⚡ Quick Answer: How to Run a Local AI Model on Your Device

🎯 The "Build for a Friend" Challenge (Hacktoberfest 2026)

📊 Hardware Sizing Matrix: What Model Can Your Computer Run?

🏗️ System Architecture

Component Roles & Boundaries

🛠️ Starter Archetypes

1. Local Chat (`hack-day-starter-chat-<model>`)

2. Autonomous Tool-Calling Agent (`hack-day-starter-agent-<model>`)

💻 Real End-to-End Execution Trace: Gemma 4 12B Agent

What Happened Behind the Scenes:

🔍 Verified Model Registry (Live Ollama Source Snapshot)

🚦 Ollama & Model Readiness UX (Zero Stack Traces)

🔒 Security & Privacy Commitments

🚀 Step-by-Step Installation & Quickstart

Method 1: Using the Web App Profiler (Recommended)

Method 2: Running Your Downloaded Starter Project

🧪 Comprehensive Automated Test Suite (36/36 Passing)

❓ Frequently Asked Questions (FAQ)

Q1: Can I run an AI model locally without a dedicated GPU?

Q2: How much RAM do I need for local AI?

Q3: Why run models locally instead of using cloud APIs like OpenAI or Anthropic?

Q4: What is the difference between Ollama and Hack Day Starter?

Q5: How does tool-calling work without an internet connection?

📁 Repository Structure

🤝 Contributing

📄 License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

🚀 Hack Day Starter — Run Local AI Models on Any Laptop or PC

📚 Technical Documentation & Guides

⚡ Quick Answer: How to Run a Local AI Model on Your Device

🎯 The "Build for a Friend" Challenge (Hacktoberfest 2026)

📊 Hardware Sizing Matrix: What Model Can Your Computer Run?

🏗️ System Architecture

Component Roles & Boundaries

🛠️ Starter Archetypes

1. Local Chat (hack-day-starter-chat-<model>)

2. Autonomous Tool-Calling Agent (hack-day-starter-agent-<model>)

💻 Real End-to-End Execution Trace: Gemma 4 12B Agent

What Happened Behind the Scenes:

🔍 Verified Model Registry (Live Ollama Source Snapshot)

🚦 Ollama & Model Readiness UX (Zero Stack Traces)

🔒 Security & Privacy Commitments

🚀 Step-by-Step Installation & Quickstart

Method 1: Using the Web App Profiler (Recommended)

Method 2: Running Your Downloaded Starter Project

🧪 Comprehensive Automated Test Suite (36/36 Passing)

❓ Frequently Asked Questions (FAQ)

Q1: Can I run an AI model locally without a dedicated GPU?

Q2: How much RAM do I need for local AI?

Q3: Why run models locally instead of using cloud APIs like OpenAI or Anthropic?

Q4: What is the difference between Ollama and Hack Day Starter?

Q5: How does tool-calling work without an internet connection?

📁 Repository Structure

🤝 Contributing

📄 License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

1. Local Chat (`hack-day-starter-chat-<model>`)

2. Autonomous Tool-Calling Agent (`hack-day-starter-agent-<model>`)