Skip to content

TOOL_STACK.md

Latest commit

 

History

History
312 lines (222 loc) · 7.73 KB

TOOL_STACK.md

File metadata and controls

312 lines (222 loc) · 7.73 KB
Raw

HeliOS Studio Tool Stack

Recommended tools for each layer of the stack, with rationale and alternatives.

Discovery Layer

Primary: Perplexity AI

Why: Best-in-class web research with citations; Collections feature for organizing research themes.

Use for:

  • Opportunity scanning ("emerging security gaps in MCP-based agents")
  • Market research ("developer pain points in AI key management 2026")
  • Competitive analysis
  • Trend identification

Alternatives:

  • ChatGPT with web search (good, but weaker citation and source quality)
  • Bing Chat (free, but less organized)

Knowledge Layer

Primary: NotebookLM

Why: Purpose-built for research; handles diverse sources (PDFs, URLs, videos); generates timelines, mind maps, audio summaries, and comprehensive Deep Research reports.

Use for:

  • Deep topic dives
  • Literature reviews
  • Experiment logs
  • Long-term project knowledge base

Features:

  • Deep Research (comprehensive multi-source analysis)
  • Multiple export formats (summaries, timelines, mind maps, flashcards, audio, video)
  • Support for 100+ sources per notebook
  • PDF, URL, Google Doc, YouTube, image ingestion

Alternatives:

  • Obsidian + local markdown (more control, less AI assistance)
  • Notion AI (better collaboration, weaker research focus)

Secondary: GitHub Knowledge Repos

Why: Version-controlled, searchable, integrates with rest of workflow.

Structure:

kb-opportunities-cybersec/
kb-opportunities-devtools/
kb-tech-rag-evals/
kb-tech-mcp-security/
kb-feeds/  # Raw RSS/API captures

Reasoning Layer

Primary: Claude (Anthropic)

Why: Strongest reasoning, longest context (200k tokens), excellent for architecture and system design.

Use for:

  • System design
  • Threat modeling
  • ADRs and RFCs
  • Architecture diagrams
  • Trade-off analysis

Access methods:

  • Web interface for quick iteration
  • Desktop app with MCP support for tool integration

Primary: Claude Code

Why: Repo-aware planning, large refactors, GitHub automation via MCP.

Use for:

  • Codebase analysis ("What's the architecture?")
  • Multi-file refactors
  • Repo scaffolding
  • GitHub automation (create repos, issues, PRs)

MCP capabilities:

  • Filesystem access (sandboxed)
  • GitHub operations
  • Custom tool integration

Secondary: Local LLMs (via Ollama)

Why: Private, offline, cost-free for suitable tasks.

Models:

  • Qwen3-7B / Llama 3-8B: General reasoning and chat
  • DeepSeek / Qwen3 thinking variants: Long reasoning chains
  • Phi-3 Mini: Fast CPU-friendly utility tasks

Use for:

  • Private/sensitive analysis
  • Offline work
  • Cheap batch processing

Coding Layer

Primary: Cursor

Why: AI-native, best-in-class repo awareness, multi-file edits, model flexibility.

Features:

  • Full repo indexing (272k+ token context)
  • Composer for multi-file edits
  • Model selection per task (Claude, GPT, Gemini, custom)
  • Tab completion
  • AI-powered debugging

Price: ~$20/month

Best for: Feature work, refactors, new projects

Secondary: VS Code + Extensions

Why: Existing workflows, extensions, corporate Copilot license.

Extensions:

  • GitHub Copilot (OpenAI models)
  • Cody (Sourcegraph, strong codebase search)
  • Continue (open-source, multi-model)

Best for: Infrastructure scripting, specialized extensions

Tertiary: Claude Code (CLI/outside IDE)

Why: Cross-repo automation, migration scripts.

Best for: Batch operations, repo setup, multi-repo changes

Local AI Layer

Primary: Ollama

Why: One-line install, 100+ models, OpenAI-compatible API, active development.

Recommended models:

Model Size Use Case Speed (est.)
Phi-3 Mini 3.8B Summarization, utility 50+ tok/s (CPU)
Qwen3 4-7B General chat, docs 30-40 tok/s (GPU)
Llama 3 8B Reasoning, coding 20-30 tok/s (GPU)
Qwen3-Coder 7B Code understanding 20-30 tok/s (GPU)
Qwen3 30B-A3B Heavy coding, analysis 10-15 tok/s (high-end GPU)

Installation:

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3
ollama pull qwen3
ollama pull phi3

API:

# OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3", "messages": [{"role": "user", "content": "Hello"}]}'

Alternatives:

  • LM Studio (GUI, easier for non-technical)
  • llama.cpp (more control, harder setup)

Automation Layer

Primary: n8n (self-hosted)

Why: Visual workflow builder, self-hosted, AI nodes, 400+ integrations.

Use for:

  • GitHub webhook → action workflows
  • Scheduled research captures
  • Experiment orchestration
  • Data pipelines

Deployment:

docker run -d \
  --name n8n \
  -p 5678:5678 \
  -v n8n_data:/home/node/.n8n \
  n8nio/n8n

Key features:

  • GitHub, GitLab, HTTP, webhook nodes
  • AI nodes (OpenAI, local models)
  • Database connectors
  • Cron scheduling
  • Error handling and retries

Alternatives:

  • Zapier (easier, not self-hosted, more expensive)
  • Airflow (more powerful, steeper learning curve)
  • Make.com (similar to Zapier)

Secondary: MCP Tools

Why: Standardized interfaces for Claude/Claude Code to call external tools.

Key servers:

  • Filesystem MCP: Sandboxed file operations
  • GitHub MCP: Repo, issue, PR management
  • n8n MCP: Trigger workflows, query status
  • Ollama MCP: Local inference

Security: Each server runs sandboxed with least-privilege access.

Infrastructure Layer

Version Control: GitHub

Why: Industry standard, excellent Actions CI/CD, free private repos, project management.

Use for:

  • Source of truth
  • Issue tracking
  • Project boards
  • CI/CD (Actions)
  • Documentation (Pages)

Containers: Docker + docker-compose

Why: Portable, reproducible, integrates with everything.

Standard stack per project:

services:
  app:
    build: .
    ports:
      - "8000:8000"
  db:
    image: postgres:16
  redis:
    image: redis:7

Virtualization: Proxmox

Why: Already in place, powerful, supports GPU passthrough.

VM allocation:

  • AI Core: Ollama, vector DBs, orchestrator APIs
  • Automation Hub: n8n, webhooks, monitoring
  • Runners: Self-hosted GitHub Actions runners
  • Prototypes: Project deployments

Cloud: Minimal VPS

Use for: Public-facing endpoints only; prefer homelab for everything else.

Providers: Hetzner, Digital Ocean, Linode (cost-effective)

Developer Environment Comparison

Feature VS Code + Copilot Cursor Windsurf JetBrains + AI
AI integration Extension-based Native Native Native
Repo awareness Limited Excellent Excellent Excellent
Multi-file edits Manual Composer Cascade Junie
Model choice OpenAI only Multi-provider Multi-provider Multi-provider
Local model support Via extensions Yes Yes Yes
Price $10/mo (Copilot) $20/mo $15/mo $10/mo
Best for Existing workflows AI-first dev Agentic flow JetBrains users

Recommendation: Start with Cursor as primary; retain VS Code for specialized tasks.

Cost Summary

Tool Cost Notes
Perplexity Pro $20/mo Optional; free tier usable
NotebookLM Free Google account required
Claude Pro $20/mo Optional; pay-per-use API available
Cursor $20/mo Best ROI for AI coding
GitHub Copilot $10/mo Already have via corp license
Ollama Free Self-hosted
n8n Free Self-hosted
Infrastructure $0-50/mo Homelab + optional VPS

Total: $50-100/month for full stack.

Next Steps

See SETUP_GUIDE.md for installation instructions.