Skip to content

STACK.md

Latest commit

 

History

History
171 lines (136 loc) · 8.8 KB

STACK.md

File metadata and controls

171 lines (136 loc) · 8.8 KB

AI Stack Manifest

Auto-referenced by the dispatcher skill. Last updated: 2026-03-22. Health: ✅ working | ⚠️ degraded | ❌ broken | 💤 not installed/inactive

Models

Model Access Cost Best For Limits Health
Claude Opus 4.6 Claude Code (self) Paid (primary) Architecture, complex reasoning, code review Rate limits apply ✅
Claude Sonnet 4.5 Claude Code (subagent) Paid Fast code gen, web research, parallel tasks Rate limits apply ✅
Claude Haiku API Cheap Classification, triage, routing Lower capability ✅
Gemini 2.5 Pro CLI (gemini) / API Free tier (250K TPM, 1K RPD) Best quality free, multimodal, 1M context 5-15 RPM ⚠️ (reinstalled 2026-03-22)
Grok (xAI) API / Web $175/mo free credits Contrarian takes, general tasks Requires data-sharing opt-in ⚠️ (verify key)
Perplexity Web Free/Pro Search-grounded research, citations No code execution ✅
Ollama (llama3) Local (ollama) Free ($0) Offline inference, privacy-sensitive tasks 4.7GB, GPU-bound ✅
SambaNova API (DevLoop) Free (20M tok/day) Highest-volume batching, R&D OpenAI-compat ✅
Mistral Large API (DevLoop) Free (1B tok/month) Code (Codestral), queued batch 2 RPM ✅
Cerebras API (DevLoop) Free (1M tok/day) Speed-critical tasks 2,600 tok/s ✅
Groq API (DevLoop) Free (~14K req/day) Real-time lint/review, whisper 30-60 RPM ✅
Mercury 2 API 10M free tokens (one-time) Fast diffusion inference OpenAI-compat ✅
OpenRouter API (DevLoop) 29 free models, 1K RPD Aggregator/fallback router Varies by model ⚠️ (verify key)

Token budget routing:

TIER 0 (local):     Ollama — $0, your GPU
TIER 1 (free API):  SambaNova → Mistral → Cerebras → Groq → Gemini
TIER 2 (cheap):     OpenRouter, Grok credits, Mercury 2
TIER 3 (expensive): Claude Code — supervised work only

NO Chinese models (ban in effect — MiniMax correlation, spam risk). DeepSeek/Kimi via OpenRouter allowed for non-sensitive summarization only when explicitly noted.

3D Reconstruction

Tool Access Cost Quality Speed Health
VGGT (Meta) Local (5GB weights needed) Free High (multi-view, not hallucination) Sub-second ⚠️ (weights not yet downloaded)
Trellis HuggingFace Space / API Free (quota) Good Moderate ✅ (DTR pipeline, 1st fallback)
Hunyuan3D-2.1 (Tencent) HuggingFace Space / Gradio API Free Good (4-view support) 95-120s ✅ (tested in DTR)
Rodin / Hyper3D fal.ai API ~$0.40/job High (quad-mesh, 4K PBR, 21K verts) Moderate ✅ (tested, in api_reconstruct_v2.py)
Tripo AI fal.ai API ~$0.40/job Good Moderate ✅ (DTR pipeline, 4th fallback)
RealityScan (Epic) Desktop / REST+gRPC API Free under $1M revenue Best-in-class (photogrammetry) Slow (batch) ✅
COLMAP Local (no-CUDA build) Free High (photogrammetry) Slow ⚠️ (failed at 360p; untested at 1080p — use with U2Net masking)
Blender Local Free N/A (mesh cleanup/import) Manual ✅
RunPod Serverless Cloud API ~$0.10/job Good (Trellis) Moderate ✅ (DTR pipeline, 2nd fallback)
GCP Spot L4 Cloud API ~$0.12/job Good (Trellis) Moderate ✅ (DTR pipeline, 3rd fallback)

DTR fallback chain: Trellis (free HF) → RunPod ($0.10) → GCP ($0.12) → Tripo ($0.40) Rule: All photogrammetry must use U2Net masking.

Voice / Audio

Tool Access Cost Notes Health
ElevenLabs API Paid (105K token budget noted) TTS, voice cloning, MC scripts, ad breaks ✅
Groq Whisper API (DevLoop / OpenWhispr) Free ~440ms latency, real-time STT ✅
OpenWhispr Local Electron app (~/openwhispr/) Free Push-to-talk, Groq-backed, RCtrl+RAlt PTT ✅
Claude Code /voice Built-in Free Improved 2026-03-15, needs headset mic ✅
openai-whisper Local pip (v20250625) Free Offline STT, slower ✅

Code / Dev

Tool Access Notes Health
Claude Code Terminal / IDE Primary coding agent (this system) ✅
Codex CLI Terminal (codex) OpenAI agent, different model perspective ✅
Cursor IDE Code-aware second opinion, inline edits ✅
Gemini CLI Terminal (gemini) Google agent, 1M context, free tier ⚠️ (reinstalled 2026-03-22, was broken)
LangGraph Python (langgraph 1.1.3) Graph-based agent workflows ✅
LiteLLM Python (litellm 1.81.6) Model gateway / provider abstraction ✅
LlamaIndex Python (llama-index 0.14.18) RAG / data indexing (activate per-project) ✅ (dormant)
ChromaDB Python (chromadb 1.5.1) Vector database ✅
E2B Sandbox — Safe agent code execution 💤 (pip install pending, $100 credit)
Arize Phoenix — Agent tracing + cost dashboard 💤 (Docker deploy pending)
Context7 MCP — Live API docs, kills hallucinated API calls 💤 (install pending)

Automation

Tool Access Notes Health
n8n Hetzner VPS (SSH tunnel) Workflow automation, SQLite, 330MiB RAM ✅
DevLoop Local (~/devloop/) LLM router + 7 free providers, 50M+ tok/day ✅ (MVP, sign-ups in progress)
GSD Claude Code skills (~/.claude/commands/gsd/) Execution framework v1.26.0 ✅
BMAD ~/_bmad/ Planning/SDLC agent framework v6.2.0 ✅
AI Hot Sauce ~/ai-hot-sauce/ (~/scrape-depot/ai-hot-sauce/) KPI library (52 KPIs), context manager, routing engine ✅
Playwright MCP — Browser automation, web scraping, E2E tests 💤 (install pending)

Research

Tool Access Notes Health
/recon skill Claude Code skill Parallel 4-agent research fan-out ✅
Perplexity Web Search-grounded, citation-heavy ✅
Gemini CLI Terminal (gemini) 1M context, free tier, web access ⚠️ (auth may need refresh)
Web search (Claude) Built-in (claude.ai) General search-grounded queries ✅

Creative

Tool Access Notes Health
Canva MCP claude.ai MCP Design gen + editing, 30+ tools ✅
Figma MCP claude.ai MCP Design context, Code Connect, FigJam diagrams ✅
Spline MCP claude.ai MCP (~/spline-mcp/) 3D code gen, WebSocket, n8n, 22 tools ✅
Remotion npm / ~/remotion/ Programmatic video creation ✅
Blender Local desktop 3D modelling, mesh cleanup, animation ✅
ElevenLabs API TTS for creative audio, MC scripts, ads ✅

Infrastructure

Provider Services Cost Health
Vercel Riffraff + DTR hosting, CI/CD Free tier ✅
Railway OpenClaw (Node.js agent gateway) Free/usage ✅
Hetzner VPS n8n, Tailscale, future OpenClaw mirror ~€4-6/mo ⚠️ (DDoS risk at L7 — CrowdSec/CF free tier TODO)
HuggingFace Model hosting, Spaces (Trellis, VGGT, Hunyuan3D) Free tier ✅
GitHub Source control, Pages (agentic taxonomy viz), agent-memory repo Free ✅
Docker Local containerisation + production deploys Free (local) ✅
Ollama Local LLM inference (llama3 4.7GB) Free ✅

MCP Servers (claude.ai)

Server Status Tools Notes
Spline MCP ✅ Active 22 3D code gen, asset mgmt, WebSocket, n8n
Figma MCP ✅ Active 10+ Design context, Code Connect
Canva MCP ✅ Active 30+ Design gen + editing
Slack MCP ✅ Active 10+ Message/channel ops
Gmail MCP ✅ Active 7 Email search/compose
Google Calendar MCP ✅ Active 9 Event mgmt, availability
Playwright MCP ✅ Active 20+ Browser automation
Notion MCP 💤 Disabled 15+ Re-enable for wiki/database work
Clay MCP 💤 Disabled 15+ Re-enable for freelance outbound sales

Lean Stack (Active Config — decided 2026-03-20)

Council decision: Stack A "Lean" + n8n. One tool per layer.

Layer Tool Status
LLM SDK Anthropic + OpenAI ✅ Active
Orchestration LangGraph (sole) ✅ Active
RAG/Data LlamaIndex (per-project only) ✅ Dormant
Vector DB ChromaDB ✅ Active
Model Gateway LiteLLM ✅ Active
Observability Arize Phoenix 💤 TODO (Docker)
Sandbox E2B 💤 TODO (pip)
MCP Docs Context7 💤 TODO
Automation n8n ✅ Live (Hetzner)
Local LLM Ollama (llama3) ✅ Active

Dropped (installed but unused): CrewAI, Vellum, LangSmith, LangChain base, Composio