⚓ Harbor

The Multi-Cloud Registry & Discovery Platform for AI Agents

Harbor is a centralized registry, discovery, and governance platform for AI agents — regardless of where they run. Agents deployed on AWS Bedrock AgentCore, Azure AI Agent Service, Google Vertex AI, or your own infrastructure all register with Harbor to be discovered, governed, and audited.

Harbor is a passport office, not an airline. Your agents fly on whatever cloud they want. Harbor issues the passport, checks compliance, and tells other agents where to find them.

The Problem

Every cloud provider offers an agent runtime. None of them answer:

Which agents exist across our organization? No central registry across clouds, accounts, and teams.
Who approved this agent for production? No lifecycle governance or approval workflow.
Can Agent A talk to Agent B? No cross-cloud communication policy enforcement.
What tools is this agent allowed to use? No capability boundary control.
What happened at 3am? No unified audit trail across your agent fleet.

These are the same problems API Management solved for microservices a decade ago. Harbor solves them for AI agents.

Why Harbor

Capability	Without Harbor	With Harbor
Agent inventory	Spreadsheets, tribal knowledge	Centralized registry with cross-cloud metadata
Deployment approval	Slack messages, hope	Lifecycle pipeline: draft → review → approve → publish
Cross-team discovery	"Hey, does anyone have an agent that does X?"	`harbor discover -c summarization`
Access control	All agents can call anything	Communication ACL + capability boundaries
Incident response	"Which agent is causing this?"	`harbor lifecycle agent-x suspended --reason "incident-1234"`
Compliance audit	Manual evidence collection	Immutable audit trail with tenant context
Multi-cloud governance	Per-cloud silos	Unified registry, one API for all providers

Key Differentiators

Cloud-agnostic registry — agents from AWS, Azure, GCP, or on-prem all register with the same API and data model
Harbor is a registry, not a deployment tool — operators deploy agents with their own tooling (Terraform, CDK, ARM Template, gcloud), then register metadata with Harbor
A2A Agent Card alignment — agent metadata follows the A2A protocol spec, extended with governance and cross-cloud fields
Hot-swappable agent lifecycle — register, publish, suspend, and retire agents without downtime; discovery always returns the latest published state
Enterprise governance — lifecycle approval pipeline with role-based access (risk officer, compliance officer sign-off for production)
Multi-tenant by design — tenant = cloud account/project/subscription; integrates with AWS Control Tower for org-wide visibility
Policy enforcement — capability boundaries, communication ACL, and schedule windows — evaluated centrally, enforced across clouds
Full observability — health monitoring, audit trail, EventBridge events, Security Hub integration

Architecture

graph TB
    subgraph Operators["Operators (deploy with their own tools)"]
        direction LR
        TF["Terraform / CDK / Pulumi"]
        ARM["ARM Template / Bicep"]
        GC["gcloud CLI"]
    end

    subgraph Runtimes["Agent Runtimes (any cloud)"]
        direction LR
        AWS_RT["AWS Bedrock<br/>AgentCore"]
        AZ_RT["Azure AI<br/>Agent Service"]
        GCP_RT["Google<br/>Vertex AI"]
        CUSTOM["On-Prem /<br/>Custom"]
    end

    subgraph Harbor["Harbor Central (AWS Serverless)"]
        CF["CloudFront + WAF"]
        APIGW["API Gateway"]
        FN["Lambda<br/>FastAPI + Mangum"]
        DB["DynamoDB<br/>Single-Table"]
        EB["EventBridge"]
        COG["Cognito"]
    end

    subgraph Consumers["Consumers"]
        direction LR
        UI["Harbor UI<br/>React SPA"]
        CLI["Harbor CLI"]
        SDK["Agent SDK /<br/>Direct API"]
    end

    TF --> AWS_RT
    ARM --> AZ_RT
    GC --> GCP_RT

    AWS_RT -- "register / heartbeat / discover" --> CF
    AZ_RT -- "register / heartbeat / discover" --> CF
    GCP_RT -- "register / heartbeat / discover" --> CF
    CUSTOM -- "register / heartbeat / discover" --> CF

    CF --> APIGW --> FN
    FN --> DB
    FN --> EB
    FN --> COG

    UI --> CF
    CLI --> CF
    SDK --> CF

    style Harbor fill:#0f172a,stroke:#3b82f6,color:#e2e8f0
    style Runtimes fill:#1e293b,stroke:#64748b,color:#e2e8f0
    style Operators fill:#1e293b,stroke:#64748b,color:#e2e8f0
    style Consumers fill:#1e293b,stroke:#64748b,color:#e2e8f0

Agent Onboarding Flow

sequenceDiagram
    participant Op as Operator
    participant Cloud as Any Cloud Runtime
    participant H as Harbor API
    participant Rev as Reviewer

    Op->>Cloud: Deploy agent (Terraform / CDK / ARM / gcloud)
    Cloud-->>Op: Agent running ✓

    Op->>H: POST /agents (metadata, runtime, endpoint, compliance)
    H-->>Op: lifecycle: draft

    Op->>H: PUT /agents/{id}/lifecycle?target=submitted
    H-->>Rev: Pending review notification

    Rev->>H: POST /reviews/{id}?action=approve
    H-->>H: lifecycle: approved → published

    Cloud->>H: PUT /agents/{id}/health (periodic heartbeat)
    Cloud->>H: GET /discover/capability/summarization
    H-->>Cloud: Best matching published agent

Agent Data Model

When registering, operators provide an "agent passport" — metadata about where the agent runs, how to reach it, and what it can do:

Section	Fields	Who Provides
Identity	`agent_id`, `name`, `description`, `version`	Operator (required)
Tenant	`tenant_id`, `owner`	Operator (required)
Runtime Origin	`provider` (aws/azure/gcp/on-prem), `runtime`, `region`, `account_id`, `resource_id`	Operator
Endpoint	`url`, `protocol` (http/grpc/a2a/mcp), `auth_type`, `health_check_path`	Operator
Capabilities	`skills`, `capabilities`, `phase_affinity`	Operator
Dependencies	`required_agents`, `required_tools`, `models`	Operator
Compliance	`data_residency`, `certifications`, `pii_handling`, `data_classification`	Operator
Governance	`lifecycle_status`, `visibility`, `routing_rules`	Harbor-managed

All fields except identity and tenant are optional with sensible defaults — low barrier to register, rich metadata for governance review.

Agent Lifecycle

stateDiagram-v2
    [*] --> Draft
    Draft --> Submitted
    Submitted --> InReview: reviewer picks up
    InReview --> Approved: passes review
    InReview --> Draft: rejected
    Approved --> Published
    Published --> Suspended: emergency kill switch
    Published --> Deprecated: sunset announced
    Suspended --> Published: reinstated
    Suspended --> Deprecated
    Deprecated --> Retired
    Deprecated --> Published: un-deprecate
    Retired --> [*]

Environment	Approval Required
dev	Auto-approve (self-publish)
staging	1 approval from project admin
prod	2 approvals: risk officer + compliance officer

Only published agents are discoverable. Suspended is an emergency kill switch any admin can trigger.

Quick Start

# Clone and setup
git clone https://github.com/yikaikao/harbor.git
cd harbor
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,cli]"

# Run locally
uvicorn harbor.main:app --reload --port 8100

# Use the CLI
export HARBOR_URL=http://localhost:8100/api/v1
export HARBOR_TENANT=dev-tenant
export HARBOR_OWNER=dev@harbor.local

# Register an agent from any cloud
harbor register my-agent "My Agent" --capabilities nlp,summarize
harbor list
harbor lifecycle my-agent submitted
harbor discover -c nlp --resolve
harbor health my-agent

See docs/getting-started.md for the full setup guide.

Runtime Policies

Capability Boundaries

Control what each agent can access:

tools:
  allowed: ["db_query", "send_email"]
  denied: ["execute_trade"]
  require_human: ["large_transfer"]
mcp_servers:
  allowed: ["internal-kb", "market-data"]
  denied: ["external-*"]
data_classification:
  max_level: "confidential"

Communication ACL

Control which agents can talk to each other (default: allowlist / deny-all):

rules:
  - from: "trading-agent"
    to: "risk-assessment-agent"
    required: true
  - from: "external-*"
    to: "internal-*"
    allowed: false

Schedule Windows

Control when agents can operate:

active_windows:
  - cron: "0 9-16 * * MON-FRI"
    timezone: "Asia/Taipei"
out_of_window_action: "reject"

Tech Stack

Layer	Technology
Frontend	React + Tailwind CSS + Vite
API	FastAPI + Mangum (Lambda adapter)
Compute	AWS Lambda (Python 3.12, ARM64)
Gateway	API Gateway HTTP API + Cognito JWT
Storage	DynamoDB (single-table, multi-tenant)
CDN	CloudFront + S3 + WAF
Auth	Cognito + IAM Identity Center
Events	EventBridge + SNS
IaC	AWS CDK (TypeScript)
CLI	Click + httpx

Project Structure

harbor/
├── src/harbor/              # Python backend (FastAPI)
│   ├── models/              # Pydantic data models (agent, policy)
│   ├── store/               # DynamoDB persistence (5 store classes)
│   │   ├── base.py          # Shared table access
│   │   ├── agent_store.py   # Agent CRUD + indexes
│   │   ├── health_store.py  # Health status
│   │   ├── audit_store.py   # Audit entries
│   │   ├── policy_store.py  # Policy CRUD
│   │   └── version_store.py # Version snapshots
│   ├── registry/            # Agent lifecycle governance
│   ├── discovery/           # Capability & phase-based lookup
│   ├── policy/              # Runtime policy enforcement
│   ├── health/              # Health monitoring & heartbeat
│   ├── audit/               # Audit log service
│   ├── auth/                # JWT validation & RBAC
│   ├── events/              # EventBridge emitter
│   ├── sync/                # A2A Agent Card import
│   ├── cli/                 # CLI tool
│   └── api/                 # FastAPI routers (6 router modules)
│       ├── deps.py          # Service container & auth
│       ├── agents.py        # Agent CRUD + lifecycle
│       ├── discovery.py     # Discovery endpoints
│       ├── health.py        # Heartbeat + summary
│       ├── audit.py         # Audit log
│       ├── policies.py      # Policy CRUD + evaluate
│       └── reviews.py       # Review queue
├── frontend/                # React SPA
├── infrastructure/          # AWS CDK
├── tests/                   # 75 Python + 18 CDK tests
└── docs/                    # Architecture, API reference, guides

Demo: Life Insurance Agent Platform

A complete sample application demonstrating Harbor + A2A + Strands Agents + Bedrock AgentCore integration. Six AI agents collaborate via A2A v1.0 protocol, managed by Harbor's lifecycle governance.

→ See full details in samples/life-insurance/README.md

Agent Registry & Discovery

$ harbor list
  explanation-agent              published    保險知識 Agent
  recommendation-agent           published    推薦引擎 Agent
  compliance-check-agent         published    合規檢查 Agent
  premium-calculator-agent       published    保費試算 Agent
  underwriting-risk-agent        published    風險預評估 Agent
  product-catalog-agent          published    商品目錄 Agent

$ harbor discover -c risk_assessment
  underwriting-risk-agent        風險預評估 Agent

Lifecycle Governance — Emergency Kill Switch

$ harbor lifecycle underwriting-risk-agent suspended --reason 'incident-1234'
✓ underwriting-risk-agent → suspended

$ harbor discover -c risk_assessment
(empty — suspended agent is no longer discoverable!)

A2A Agent Collaboration on AgentCore Runtime

The Orchestrator agent coordinates multiple specialist agents (product search, premium calculation, risk assessment) to produce a complete insurance plan:

>>> Invoking Orchestrator Agent via A2A protocol on AgentCore Runtime...
>>> Prompt: 我30歲，月預算3000元，想買醫療險，請幫我搜尋商品並試算保費

<<< Agent Response (26.0s):

## 搜尋結果與保費試算

### 方案一：國泰真全意住院醫療險 ⭐推薦
- 月保費：833元（住院日額 2,000元 / 手術限額 20萬元）

### 方案二：富邦人生自由行醫療險
- 月保費：977元（住院日額 2,500元 / 手術限額 25萬元）

### 風險評估結果
核保風險等級：「優體」，兩款商品都有15%的折扣。

Frontend — Agent Status Dashboard

Documentation

Architecture — system design, data model, lifecycle, policies
Getting Started — local dev setup, API & CLI usage
API Reference — all 29 endpoints with examples
Enterprise Integration Guide — Control Tower deployment runbook
IAM Identity Center Setup — SSO configuration

API Overview

29 REST endpoints under /api/v1:

Category	Endpoints
Agent CRUD	POST/GET/PATCH/DELETE `/agents`
Lifecycle	PUT `/agents/{id}/lifecycle`
Versions	POST/GET `/agents/{id}/versions`
Health	PUT `/agents/{id}/health`, GET `/health/summary`
Audit	GET `/agents/{id}/audit`, GET `/audit`
Discovery	GET `/discover/capability/{cap}`, `/discover/phase/{phase}`, `/discover/resolve`
Policies	CRUD for capability, communication, schedule policies
Reviews	GET `/reviews/pending`, POST `/reviews/{id}`

See docs/api-reference.md for full request/response examples.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 12 Commits
.kiro		.kiro
design-system/harbor		design-system/harbor
docs		docs
frontend		frontend
infrastructure		infrastructure
samples/life-insurance		samples/life-insurance
scripts		scripts
src/harbor		src/harbor
tests		tests
.gitignore		.gitignore
PLAN.md		PLAN.md
README.md		README.md
pyproject.toml		pyproject.toml

Navigation Menu

Repository files navigation

⚓ Harbor

The Problem

Why Harbor

Key Differentiators

Architecture

Agent Onboarding Flow

Agent Data Model

Agent Lifecycle

Quick Start

Runtime Policies

Capability Boundaries

Communication ACL

Schedule Windows

Tech Stack

Project Structure

Demo: Life Insurance Agent Platform

Agent Registry & Discovery

Lifecycle Governance — Emergency Kill Switch

A2A Agent Collaboration on AgentCore Runtime

Frontend — Agent Status Dashboard

Documentation

API Overview

License

About

Releases

Packages

Used by

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

⚓ Harbor

The Problem

Why Harbor

Key Differentiators

Architecture

Agent Onboarding Flow

Agent Data Model

Agent Lifecycle

Quick Start

Runtime Policies

Capability Boundaries

Communication ACL

Schedule Windows

Tech Stack

Project Structure

Demo: Life Insurance Agent Platform

Agent Registry & Discovery

Lifecycle Governance — Emergency Kill Switch

A2A Agent Collaboration on AgentCore Runtime

Frontend — Agent Status Dashboard

Documentation

API Overview

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages