Skip to content

raphaelmansuy/edgecore

Repository files navigation

███████╗██████╗  ██████╗ ███████╗ ██████╗ ██████╗ ██████╗ ███████╗
██╔════╝██╔══██╗██╔════╝ ██╔════╝██╔════╝██╔═══██╗██╔══██╗██╔════╝
█████╗  ██║  ██║██║  ███╗█████╗  ██║     ██║   ██║██████╔╝█████╗
██╔══╝  ██║  ██║██║   ██║██╔══╝  ██║     ██║   ██║██╔══██╗██╔══╝
███████╗██████╔╝╚██████╔╝███████╗╚██████╗╚██████╔╝██║  ██║███████╗
╚══════╝╚═════╝  ╚═════╝ ╚══════╝ ╚═════╝ ╚═════╝ ╚═╝  ╚═╝╚══════╝

The Open-Source, Self-Hosted AI Agent Platform for Kubernetes

The Kubernetes of AI Agents — framework-agnostic, MicroVM-isolated, data-sovereign

License: Apache 2.0 Go Version Kubernetes MCP A2A Kata Containers


Why EdgeCore Exists

In 2026, AI agents run production workflows. They answer questions, write code, orchestrate business processes. And yet — every serious production deployment is locked inside a vendor's cloud.

AWS Bedrock AgentCore. Google Gemini Enterprise Agent Platform. Azure AI Foundry. Excellent products. But you do not control them. Your data lives on their infrastructure, under their policies, at their prices.

This is the same problem open source solved before:

Proprietary era         Open-source disruption       New status quo
─────────────────       ──────────────────────        ────────────────
Oracle              →   PostgreSQL               →   Postgres everywhere
Windows Server      →   Linux                    →   Linux everywhere
VMware              →   KVM + containers         →   Kubernetes everywhere
AWS Bedrock         →   EdgeCore                 →   Agents everywhere

EdgeCore is the Kubernetes of AI agents — open, self-hosted, Kubernetes-native, Apache 2.0. No rug-pull possible.

Read the full rationale: MANIFESTO.md · Why & Vision · Problem Space


What EdgeCore Does

EdgeCore is a production-grade AI Agent Platform that matches the capabilities of AWS AgentCore and Google's Agent Engine while adding what managed clouds cannot offer:

Capability EdgeCore AWS AgentCore Google Agent Engine
Vendor independence
Air-gapped / on-prem
MicroVM hardware isolation
Framework agnostic Partial ADK-preferred
Full data sovereignty
Apache 2.0 license
Kubernetes-native CRDs
Custom isolation model

See the full gap analysis: SOTA Assessment (May 2026)


Architecture: 7 Layers

External World (Users, Other AI Systems, Event Sources)
         |
         v
+─────────────────────────────────+
│  Layer 1: Ingress               │  Istio Gateway + JWT validation + mTLS
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 2: Control Plane         │  EdgeCore Operator (Kubebuilder v4.4)
│                                 │  AgentDeployment / AgentSession /
│                                 │  AgentPolicy / AgentEvalJob CRDs
│                                 │  Kagent MCP+A2A Gateway · Keycloak IdP
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 3: Compute Plane         │  Kata Containers + Firecracker MicroVMs
│                                 │  One kernel per agent (hardware isolation)
│                                 │  KEDA scale-to-zero autoscaling
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 4: Agent Container       │  Any framework: ADK 2.0 / LangGraph /
│                                 │  CrewAI / Strands / PydanticAI / Custom
│                                 │  Contract: GET /health + OTEL export
│                                 │  Dapr sidecar for state + pub/sub
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 5: Service Mesh          │  Istio mTLS · OPA policy on tool calls
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 6: LLM Gateway           │  LiteLLM Proxy (100+ providers)
│                                 │  ModelCatalogEntry CRD · cost tracking
+─────────────────────────────────+
         |
         v
+─────────────────────────────────+
│  Layer 7: Platform Services     │  Dapr + Redis / PostgreSQL
│                                 │  JuiceFS + MinIO · pgvector
│                                 │  External Secrets Operator + Vault
│                                 │  OTel + LGTM stack · AgentEvalJob
+─────────────────────────────────+

Deep dive: Architecture Overview · Compute Plane · Security & Identity


Core Features

MicroVM Isolation

Every agent runs in a Kata+Firecracker MicroVM with its own Linux kernel. Container escapes cannot cross agent boundaries. Cold start in 150–300 ms.

Compute Plane documentation

Framework Agnosticism

The platform contract is minimal: a container image, GET /health, and OTEL export. Build agents with any framework:

Framework Version Stars Notes
Google ADK v2.0 Beta (Py) / v1.0 (TS/Go/Java) 19.4k Graph workflows, Ambient Agents, A2A native
LangGraph v1.1 stable 30.9k Durable execution, checkpointers, HITL
CrewAI v1.14 50.4k Multi-agent flows, built-in guardrails
OpenAI Agents SDK v0.14.8 12k+ Sandbox agents, voice pipeline, handoffs
PydanticAI v1.89+ Structured outputs, type-safe tools
Strands v0.x 6.3k Model-driven, MCP native
Dapr Agents v1.14 24k Actor model, Dapr bindings

Agent Frameworks Guide

Native Protocol Support

  • MCP v1 (Model Context Protocol) — universal tool calling, adopted by Anthropic, OpenAI, GitHub, VS Code
  • A2A v1.0 (Agent-to-Agent) — Linux Foundation standard, Google/AWS/MS implementing

Gateway & Protocols

True Scale-to-Zero

KEDA-managed deployments with minReplicas=0. Agents sleep between tasks. Warm pool delivers sub-50ms wakeup.

Four-Layer Security

Layer 4: Semantic Guardrails (prompt injection detection, Llama Guard)
Layer 3: Identity (Keycloak OIDC, SPIFFE JWT-SVIDs, Kyverno admission)
Layer 2: Policy (OPA/Rego — policy-as-code, default deny on all tool calls)
Layer 1: Isolation (Kata+Firecracker MicroVM — one kernel per agent)

Security & Identity · Delegated Auth

Evaluation as CI

AgentEvalJob CRD runs quality gates on every deployment. LLM-as-judge. Golden datasets. Regression prevention. Integrates with Braintrust, Phoenix/Arize, and PromptFoo.

Evaluation Framework


Quick Start (Local, macOS/Linux)

Prerequisites

  • macOS: OrbStack with K3S enabled — OR — Linux with K3S installed
  • Docker
  • kubectl
  • go 1.24+

1. Install CRDs

kubectl apply -f operator/config/crd/bases
kubectl apply -f operator/config/namespaces
kubectl apply -f operator/config/rbac

2. Build and run the operator

cd operator
make build
make run

3. Deploy your first agent

kubectl apply -f examples/hello-agent/deploy.yaml
kubectl get agentdeployments -A

4. Run the examples locally (no cluster needed)

# Invoice classifier (PydanticAI + Ollama)
cd examples/invoice-classifier
pip install -r requirements.txt
python server.py

# Order router (Google ADK multi-agent + Ollama)
cd examples/order-router
pip install -r requirements.txt
python server.py

# Open WebUI AP assistant (connects to Open WebUI chat UI)
cd examples/openwebui-agent
pip install -r requirements.txt
python server.py

Full walkthrough: Tutorial 1 — Bootstrap · Tutorial 2 — First Agent · Local Dev Guide


Examples

Example Framework Description
hello-agent Go (stdlib) Minimal agent: health endpoint, LLM gateway, context compression, OTEL
invoice-classifier PydanticAI 1.89 + Ollama Auto-classifies invoices, extracts line items, flags anomalies — structured output
order-router Google ADK + Ollama (gemma4) Multi-agent: orchestrator + inventory + shipping sub-agents; routes orders to optimal warehouse
openwebui-agent PydanticAI + SSE AP assistant connecting directly to Open WebUI via OpenAI Chat Completions API

Each example ships with a Dockerfile, deploy.yaml (AgentDeployment CRD), and a test_*.py test suite.


CRDs at a Glance

EdgeCore's control plane is entirely CRD-driven. No proprietary SDKs.

# AgentDeployment — the core primitive
apiVersion: edgecore.io/v1
kind: AgentDeployment
metadata:
  name: invoice-classifier
  namespace: agents
spec:
  image: edgecore/invoice-classifier:1.0.0
  framework: pydantic-ai
  compute:
    runtime: kata-fc         # kata-fc | gvisor | runc
    cpu: "500m"
    memory: "512Mi"
  scaling:
    minReplicas: 0           # scale-to-zero
    maxReplicas: 10
  hitl:
    enabled: true            # human-in-the-loop approval
    approvalTimeoutSeconds: 300
  artifacts:
    enabled: true
    retentionDays: 30
  llmGateway: true           # route all LLM calls through LiteLLM proxy
  policyRef:
    name: invoice-policy

All CRDs:

CRD Purpose
AgentDeployment Lifecycle + compute + scaling for one agent
AgentSession Per-user session with TTL and pod binding
AgentPolicy OPA-backed tool and action policies
AgentCatalogEntry Discovery registry with skill-based semantic search
AgentEvalJob CI quality gate: golden datasets + LLM-as-judge
AgentExpositionBinding Surface agent to UI/API/A2A with access control
AgentAccessPolicy Who can invoke which agent
ModelCatalogEntry LLM models registered in the gateway
ToolRegistry / CLIToolRegistry MCP tool sets available to agents
AmbientAgentTrigger Event-driven triggers (Kafka/NATS/webhook)
CodingAgentDeployment Wrap CLI agents (Claude Code, Codex, Hermes)
AgentWorkspace POSIX workspace (JuiceFS/tmpfs) for coding agents
DelegatedCredential Short-lived RFC 8693 token exchange per task

Bring Your Own Agent (BYOA)

EdgeCore is a first-class host for external agents you already use:

Category Agents Platform primitive
CLI / Server agents Claude Code, OpenAI Codex CLI, GitHub Copilot CLI, Pi, Hermes Agent, OpenClaw CodingAgentDeployment + AgentWorkspace CRDs
Local LLM inference Llama 3.x, Mistral, Qwen, Phi (via vLLM / Ollama) ModelCatalogEntry registered in LLM Gateway
External API agents Any HTTP agent with OAuth or API key auth AgentProxyBinding CRD

Every BYOA agent gets: MicroVM isolation, secret injection, OTel observability, OPA policy enforcement, A2A interoperability, and catalog registration.

BYOA Guide


Roadmap: Six Phases

Phase 1: Core Operator + MicroVM Isolation         ✅ IMPLEMENTED
  Kubebuilder operator · AgentDeployment/Session CRDs
  Kata+Firecracker runtime · Basic RBAC

Phase 2: State + SOTA Parity                       🔨 DESIGNED
  Dapr state + pub/sub · KEDA scale-to-zero
  LiteLLM LLM Gateway · Ambient Agent Triggers
  HITL first-class · Context Compression

Phase 3: Tool Governance + Quality                 📋 DESIGNED
  Kagent MCP Gateway · OPA/Rego enforcement
  AgentEvalJob (evals-as-CI) · Multi-modal (voice/vision)

Phase 4: Identity + WASM                           📋 DESIGNED
  Keycloak + SPIFFE full integration
  RFC 8693 Token Exchange · WASM stateless tools

Phase 5: Scale + BYOA                              📋 DESIGNED
  JuiceFS file-based state · MicroVM snapshots
  CodingAgentDeployment · AgentWorkspace · Quotas

Phase 6: Full Observability + A2A Orchestration    📋 DESIGNED
  LGTM stack (Loki/Grafana/Tempo/Mimir)
  Quality alerting · Multi-agent orchestration patterns

Every phase is independently deployable and produces a working system. Each phase reduces a specific risk class.

Full details: Roadmap · Technical Implementation Plans · SOTA Assessment


Tutorials

# Tutorial What You Build
1 Bootstrap the Platform Full local stack: CRDs, operator, platform services
2 Deploy Your First Agent AgentDeployment → operator → pod → eval → catalog
3 Add Governance & Identity HITL approval, tool restrictions, API keys, token exchange
4 Observe, Debug, Iterate Live reconciliation, persistence, troubleshooting
5 Multi-Agent Orchestration Orchestrator + specialist pattern, A2A delegation (ADK order-router)
6 Bring Your Own Agent Wrap Claude Code or Codex CLI into the governed platform
7 Evaluation as CI Golden datasets, LLM-as-judge, pytest CI integration
8 Scale, State, Persistence Scale-to-zero proof, workspace persistence, context compression
9 Advanced Security Four-layer security: Kata + Istio mTLS + OPA + Content Guardrails
10 CLI Tools & Automation CLIToolRegistry, AmbientAgentTrigger, event-driven CI

Documentation Map

MANIFESTO.md          ← Why EdgeCore exists (start here)
docs/
  00-sota-assessment  ← Honest gap analysis vs AWS/Google/Azure (May 2026)
  01-why-and-vision   ← First-principles rationale
  02-problem-space    ← The 7 hard problems of agent infrastructure
  03-architecture     ← 7-layer system design, all CRDs, data flows
  04-agent-frameworks ← ADK 2.0, LangGraph, CrewAI, OpenAI SDK, PydanticAI
  05-compute-plane    ← Kata+Firecracker MicroVMs, KEDA, WASM
  06-gateway-and-protocols ← MCP v1, A2A v1.0, Kagent, CLI-as-MCP
  07-memory-and-state ← Dapr, Redis, JuiceFS, pgvector, Artifacts API
  08-security-and-identity ← 4-layer security pyramid, Keycloak, OPA
  09-observability    ← LGTM stack, agent tracing, quality gating
  10-roadmap          ← Six phases, SOTA gap closure plan
  11-cli-tools        ← CLI tools as first-class CRD citizens
  12-agent-exposition ← OpenWebUI, developer API, multi-tenant routing
  13-agent-catalog    ← AgentCatalogEntry, skill-based discovery
  14-delegated-auth   ← RFC 8693 token exchange, Transaction Tokens
  15-local-dev        ← OrbStack+K3S macOS, Tilt hot-reload, dev tiers
  16-evaluation       ← AgentEvalJob, LLM-as-judge, evals-as-CI
  17-multimodal       ← Voice STT/LLM/TTS, Gemini Live, OpenAI Realtime
  18-byoa             ← Claude Code, Codex CLI, Hermes Agent, OpenClaw
technical/
  00-why-first-principles  ← The constraints that govern ALL phases
  01-phase-1-core-operator ← Kubebuilder + MicroVM (8 weeks)
  02-phase-2-state-parity  ← Dapr + KEDA + SOTA G1–G4 (8 weeks)
  03-phase-3-gateway-quality ← MCP + OPA + Evals + Multimodal (8 weeks)
  04-phase-4-identity      ← Keycloak + SPIFFE + TTS + WASM (4 weeks)
  05-phase-5-scale-state   ← JuiceFS + BYOA + Snapshots (6 weeks)
  06-phase-6-observability ← LGTM + Quality Alerts + A2A (8 weeks)
tutorial/              ← Step-by-step tutorials (10 sessions)
examples/              ← Working agent code (4 agents)
operator/              ← Go Kubebuilder operator (Phase 1 implementation)

Local Development

EdgeCore supports three dev tiers — no cloud account required:

Tier Stack RAM Use case
Tier 1 Docker only 2 GB Fastest inner loop, no Kubernetes
Tier 2 OrbStack + K3S (macOS) 4 GB Full CRD/operator experience, runc runtime
Tier 3 K3S + Istio ambient 8 GB Production-parity concepts, mTLS

edgecore dev up starts the full local stack. Tilt hot-reload propagates code changes in under 5 seconds.

Local Development Guide


Security

EdgeCore takes security seriously at four independent layers:

  1. MicroVM — Kata+Firecracker gives each agent its own kernel; container escapes cannot cross boundaries
  2. mTLS + OPA — Istio encrypts all inter-service traffic; OPA enforces policy-as-code on every tool call
  3. Identity — Keycloak OIDC + SPIFFE JWT-SVIDs; RFC 8693 token exchange prevents confused deputy attacks
  4. Semantic Guardrails — Llama Guard detects prompt injection and unsafe content at every I/O boundary

Secrets are never baked into images. External Secrets Operator injects them as tmpfs-mounted env vars.

Security & Identity · Delegated Auth · Tutorial 9: Advanced Security


Key Technologies

Layer Technology Version Purpose
Control plane Kubebuilder v4.4 CRD-based operator framework
Isolation Kata Containers + Firecracker prod MicroVM per agent
Autoscaling KEDA v2.15 Scale-to-zero on any trigger
State Dapr v1.14 Backend-agnostic state + pub/sub
Identity Keycloak + SPIFFE OIDC + workload identity
Policy OPA / Rego Policy-as-code, default deny
LLM Gateway LiteLLM Proxy 100+ providers, cost tracking
Observability OpenTelemetry + LGTM Traces, metrics, logs, dashboards
Service mesh Istio mTLS, traffic policy
Files JuiceFS + MinIO Distributed POSIX filesystem
Vector store pgvector Semantic memory
Secrets External Secrets Operator Vault / AWS SM / GCP SM
Protocols MCP v1 + A2A v1.0 stable Tool calling + agent federation

Project Structure

edgecore/
├── MANIFESTO.md          # Why EdgeCore exists
├── Makefile              # Top-level build + test targets
├── operator/             # Go Kubebuilder operator (Phase 1)
│   ├── api/v1/           # CRD type definitions
│   ├── cmd/              # Operator + HITL server entry points
│   ├── config/           # Kustomize manifests (CRDs, RBAC, etc.)
│   ├── internal/         # Controller, reconciler, HITL logic
│   └── go.mod
├── examples/
│   ├── hello-agent/      # Minimal Go agent (reference implementation)
│   ├── invoice-classifier/  # PydanticAI structured extraction
│   ├── order-router/     # Google ADK multi-agent system
│   └── openwebui-agent/  # Open WebUI chat UI integration
├── docs/                 # 18 design + architecture documents
├── technical/            # 6-phase implementation plans (HOW to build)
├── tutorial/             # 10 step-by-step tutorials (HOW to use)
└── scripts/
    └── validate-docs.py  # Documentation validation CI

Build & Test

# Run all tests
make test

# Test the operator (Go)
make test-operator

# Test HITL logic
make test-hitl

# Validate CRD manifests (dry run against cluster)
make test-manifests

# Build the operator binary
make build-operator

Contributing

EdgeCore is Apache 2.0. Community-owned. No rug-pull possible.

Contributions welcome:

  • Operator features — implement Phase 2–6 capabilities from the roadmap and technical plans
  • Agent examples — add new framework examples under examples/
  • Documentation — improve tutorials or design docs
  • Integrations — new MCP tools, A2A adapters, framework integrations

See the SOTA Assessment for the seven open gaps (G1–G7) that are the highest-priority contributions.


License

Apache License 2.0 — free forever, community-owned.


Further Reading

Topic Document
Vision and first principles Why & Vision
Honest competitive analysis SOTA Assessment
Full architecture Architecture Overview
Six delivery phases Roadmap
MicroVM isolation Compute Plane
MCP + A2A protocols Gateway & Protocols
Memory, state, context Memory & State
Security pyramid Security & Identity
Traces and metrics Observability
Voice + vision agents Multi-modal
Claude Code, Codex, Hermes Bring Your Own Agent
Local developer experience Local Development
Quality gates in CI Evaluation Framework
RFC 8693 token delegation Delegated Authentication
Manifesto MANIFESTO.md

Open · Self-Hosted · Kubernetes-Native · AI Agent Platform

Apache 2.0 — free forever — community owned

About

Open-source, Kubernetes-native, framework-agnostic AI Agent Platform — the self-hosted equivalent of AWS Bedrock AgentCore / Google Agent Engine. Apache 2.0.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages