IMPORTANT: This repository contains real, production-ready, battle-tested code extracted directly from active commercial systems (like Agency OS or Founder Growth OS), rather than simplified mock learning artifacts.
For project walkthroughs, architecture flowcharts, and system context, visit the live landing page: my-portfolio-github-io-beta-five.vercel.app/projects/equilibrium.html
Building the nervous system of autonomous software.
AgentKernel is a production-first infrastructure layer for autonomous systems.
It provides the core capabilities required to build reliable AI-powered software:
- Orchestration
- Memory
- Routing
- Execution
- Recovery
- Observability
Most AI frameworks help agents think.
AgentKernel helps autonomous systems operate.
Six engines — how they compose:
graph LR
APP([Your App]) --> R01
subgraph INFRA ["AgentKernel Engines"]
direction TB
R01[01 Router\nLLM routing + circuit breaker\nBedrock → OpenAI → Ollama]
R02[02 Memory\nSHA-256 idempotency cache\nSCAR repeat-failure guard]
R03[03 Retriever\nWeb search + Firecrawl\ndependency graphing]
R04[04 Queue\nRedis distributed queue\nSSE streaming + concurrency]
R05[05 Media\nTTS 6 providers\nRemotion video rendering]
R06[06 Auth\nJWT + multi-tenant\nPrisma / SQLAlchemy]
end
R01 --> R02
R02 --> R03
R03 --> R04
R04 --> OUT([Artifact])
R05 --> OUT
R06 --> OUT
style APP fill:#0f172a,stroke:#6366f1,color:#818cf8
style R01 fill:#1e293b,stroke:#6366f1,color:#f8fafc
style R02 fill:#1e293b,stroke:#818cf8,color:#f8fafc
style R03 fill:#1e293b,stroke:#a855f7,color:#f8fafc
style R04 fill:#1e293b,stroke:#a855f7,color:#f8fafc
style R05 fill:#1e293b,stroke:#f59e0b,color:#f8fafc
style R06 fill:#1e293b,stroke:#10b981,color:#f8fafc
style OUT fill:#0f172a,stroke:#10b981,color:#10b981
Architecture reference (all 6 engines annotated):
Animated engine map (open in browser):
The visual/ folder contains visual-agentkernel.html — a standalone animated diagram showing all 6 engines lighting up as a request flows through the stack. No dependencies, no build step.
open visual/visual-agentkernel.html
# or: python -m http.server 8080 → localhost:8080/visual/visual-agentkernel.html
Full portfolio case study with live animations: my-portfolio-github-io-beta-five.vercel.app/projects/equilibrium.html
AgentKernel is six modular, production-ready engines written in both Async Python and ESM JavaScript. Every engine is independently useful; use one or wire them together.
| Engine | What It Does | Location |
|---|---|---|
| 01 Router | Multi-provider LLM routing with circuit breakers, fallover chains, token optimization | python/engines/01_router/, esm/engines/01_router/ |
| 02 Memory | Sovereign cached memory (SCAR repeat-failure guard, SHA-256 idempotency cache) | python/engines/02_memory/, esm/engines/02_memory/ |
| 03 Retriever | Web search, Firecrawl scraping, dependency graphing, content analysis | python/engines/03_retriever/, esm/engines/03_retriever/ |
| 04 Queue | Distributed task queue with circuit breaker, concurrency control, SSE streaming | python/engines/04_queue/, esm/engines/04_queue/ |
| 05 Media | TTS voice synthesis (6 providers), subtitle generation, story templates | python/engines/05_media/, esm/engines/05_media/ |
| 06 Auth | JWT auth, multi-tenant database CRUD, Prisma/SQLAlchemy schemas | python/engines/06_auth/, esm/engines/06_auth/ |
New to AgentKernel? Start here. Choose based on what you're building:
Best for: Testing LLM routing, learning fallback chains, optimizing tokens
cd python/engines/01_router
python -m venv venv && source venv/bin/activate
pip install httpx pyjwt
# Edit .env with your API keys (or just use Ollama)
python -c "from router import Router; r = Router(); print('Router ready!')"Time: 5 minutes | Cost: $0 (if using Ollama)
Best for: Building research agents, RAG systems, content analysis pipelines
cd python
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
python examples/research-video-assistant/run_assistant.pyTime: 15 minutes | Cost: $0-2 (search API free tier)
Best for: Production AI apps, outreach automation, end-to-end pipelines
docker-compose up --build
# Full stack running at http://localhost:8000
# Python API, Redis queue, SQLite, video rendering, authTime: 30 minutes | Cost: $0-5 (with Docker, no external APIs)
| Your Problem | AgentKernel Solution |
|---|---|
| "LLM costs blowing up" | Engine 01: token optimizer saves 20-30%, fallback chains = no vendor lock-in |
| "API goes down = app down" | Every engine has keyless fallback (Ollama, DuckDuckGo, in-memory) |
| "Can't deploy reliably" | Docker-compose.yml provided, runs locally or any cloud |
| "Search APIs are expensive" | Engine 03: $0 DuckDuckGo fallback if paid APIs fail |
| "Don't know how to do video" | Engine 05: Remotion template + TTS, render in minutes |
| "Auth is annoying" | Engine 06: JWT + PBKDF2, zero external deps, just works |
| "Same error keeps happening" | Engine 02: SCAR guard blocks repeated failures |
| "Too many moving parts" | All engines work standalone OR together — pick what you need |
- Failover Chains: Automated transitions across OpenAI, Anthropic, Gemini, and local Ollama channels.
- Model Prompts Styling: Automatically wraps inputs into XML tags for Gemini, constraints for Moonshot, and concise frames for Nova.
- Security Checkpoints: Active input prompt injection blockades and key output redaction streams.
- SHA-256 Response Cache: Automatically avoids duplicate LLM invocation costs.
- SCAR (Sovereign Critical Action Record): Tracks incident errors. If the same fingerprint is detected twice, it injects a warning
STOPblock directly at the top of the prompt payload.
- Graphify AST: Parses folder files recursively, outputting dependency nodes and blast-radius vectors using Python's
astpackage and JS regex parsers. - Aggregated Search & Keyless Fallback: Supports Tavily, SerpAPI, and Brave search channels, falling back to a keyless DuckDuckGo scraper if API credentials are missing.
- Firecrawl Scraper: Performs structured lead extraction from targets (e.g., IndiaMART product sheets).
- Redis Queue Manager: Distributed worker heartbeats and priority execution queues, falling back to an in-memory event queue when Redis is unavailable.
- Event Streaming: Server-Sent Events (SSE) server stream handlers with built-in connection ping pings.
- Circuit Breaker: Standalone async circuit breakers to prevent infinite execution loops and cascade network blocks.
- Universal Remotion template: Vertically aligned React composition featuring Ken Burns image pan zooms, audio synchronizers, captions overlay, and takeaway moral cards.
- Voice synthesizers: Handles ElevenLabs voice synthesis and Gemini prebuilt TTS voices, compiling raw PCM formats into WAV containers.
- ChaiPitch copywriter: AI messaging system generating Hinglish WhatsApp pitches for Indian D2C leads.
- Native Security: Native crypto-based PBKDF2 password hashing and base64 JWT token signatures without external dependencies.
- Entity CRUD: Prepared schemas and CRUD operations for Users, Leads, and Outreach messages.
agentkernel/
├── LICENSE ← MIT License
├── README.md ← Recruiter overview
├── QUICK_START.md ← Fast setup guide
├── ARCHITECTURE.md ← In-depth design patterns
├── docker-compose.yml ← Container services
│
├── python/ ← Python Suite
│ ├── engines/
│ │ ├── 01_router/ ← LLM Router & Guardrails
│ │ ├── 02_memory/ ← Cache & SCAR Guard
│ │ ├── 03_retriever/ ← Graphify & Scrapers
│ │ ├── 04_queue/ ← Redis Queue & Circuit Breakers
│ │ ├── 05_outreach/ ← ChaiPitch copywriter
│ │ └── 06_auth/ ← Database & Hashing
│ ├── pyproject.toml
│ └── requirements.txt
│
└── esm/ ← ESM JavaScript Suite (ESModules)
├── engines/
│ ├── 01_router/ ← LLM Router & Guardrails
│ ├── 02_memory/ ← Cache & SCAR Guard
│ ├── 03_retriever/ ← Graphify & Scrapers
│ ├── 04_queue/ ← Redis Queue & Circuit Breakers
│ ├── 05_media/ ← Remotion, Voice, ChaiPitch JS
│ └── 06_auth/ ← Database & Prisma config
├── package.json
└── tsconfig.json
The examples/research-video-assistant/ folder shows all 6 engines wired together into a single app: scrape leads, store them, generate Hinglish outreach copy, route through the LLM, cache the response, and render a Remotion video with TTS voice.
graph TD
A[Scraper & Retriever Engine 03] -->|Raw Lead Data| B(Auth & Database Engine 06)
B -->|Persisted Leads| C[ChaiPitch Outreach Engine 05]
C -->|Draft Message Prompts| D[Multi-Provider LLM Router Engine 01]
D -->|Active Input Guardrails Check| D1[Guardrails & Sanitizer]
D1 -->|Token-Optimized Prompts| E[Model Failover Chain]
E -->|API Timeout or Error| E1[Fallback Provider]
E -->|Successful Generation| F[Sovereign Cache Memory Engine 02]
F -->|Fingerprinted SHA-256 Hit| E
F -->|SCAR Repeat Failures Guard| D
E -->|Narration Script & Copy| G[Remotion Video Composition Engine 05]
G -->|ElevenLabs/Gemini TTS Voice| H[Video rendering output]
style A fill:#4CAF50,stroke:#333,stroke-width:2px,color:#fff
style D fill:#2196F3,stroke:#333,stroke-width:2px,color:#fff
style F fill:#9C27B0,stroke:#333,stroke-width:2px,color:#fff
style G fill:#E91E63,stroke:#333,stroke-width:2px,color:#fff
Run it:
# Python
python examples/research-video-assistant/run_assistant.py
# JavaScript
node examples/research-video-assistant/run_assistant.js