What is Agentcy?

Agentcy is an AI-powered agent orchestration platform that connects to your infrastructure, builds a knowledge graph of everything, and gives you a conversational AI that can see and act across all your systems — from a single interface.

The Problem

Modern engineering teams rely on dozens of tools: GitHub for code, AWS for compute, Kubernetes for orchestration, PostgreSQL for data, Vercel for deploys, and on and on. Each tool has its own dashboard, its own mental model, and its own API. When an incident fires at 3 AM, you flip between 10 tabs trying to piece together what went wrong. When a new engineer joins, they spend weeks just mapping out how things connect.

There is no single place to see, search, and act on your entire stack.

The Solution

Agentcy ingests data from all your systems into a unified knowledge graph (powered by Neo4j), then exposes an AI agent that can:

• Search across repositories, cloud resources, database schemas, deployments, and teams in one query
• Visualize the relationships between services, repos, clusters, and infrastructure
• Execute live actions on any connected system through real-time connector tools
• Orchestrate multi-step workflows with sub-agents, triggers, and schedules
• Enforce zero-trust policies on every action with OPA/Rego and full audit logging

Architecture at a Glance

┌───────────────────────────────────────────────────────────┐
│                     Frontend (Next.js 16)                 │
│   Chat  ·  Graph Explorer  ·  Connectors  ·  Workflows   │
└──────────────────────────┬────────────────────────────────┘
                           │ REST / SSE
┌──────────────────────────▼────────────────────────────────┐
│                  Backend API (Rust / Axum)                 │
│  agentcy-chat · agentcy-rag · agentcy-ingest · agentcy-auth · agentcy-policy       │
├───────────┬───────────┬───────────┬───────────────────────┤
│ PostgreSQL│   Neo4j   │   Redis   │  OpenFang (optional)  │
│   (SQL)   │  (Graph)  │ (Queue)   │  (Sub-agents)         │
└───────────┴───────────┴───────────┴───────────────────────┘

Key Capabilities

Universal Data Connectors

Connect to 15+ data sources, each implementing a dual-trait architecture: IngestionSource for ETL into the knowledge graph, and ConnectorToolProvider for live, real-time tool execution by the AI agent.

Category	Connectors	Live Tools
Code & DevOps	GitHub (PAT, OAuth, App)	12 tools (repos, PRs, issues, commits, files, actions)
Cloud	AWS, GCP	8 tools each (EC2, S3, Lambda, IAM / GCE, GCS, GKE, IAM)
Platform	Vercel, Supabase	8-10 tools each (deployments, projects, domains / tables, auth, storage)
Databases	PostgreSQL, MySQL, SQL Server, MongoDB	8-10 tools each (query, schema, stats, indexes)
Infrastructure	Kubernetes	10 tools (pods, deployments, services, logs, namespaces)
API	OpenAPI, MCP (Model Context Protocol)	Dynamic (auto-discovered from spec / server)
Files	CSV, JSON	Schema inference, bulk import
Execution	Remote Execution	Command execution on remote hosts
BI	Power BI	Dataset queries, report metadata

Knowledge Graph

All ingested data lives in a Neo4j 5 graph database with:

• Full-text search across all nodes and properties
• Vector embeddings for semantic search via fastembed (local, no API calls)
• Interactive visualization — explore your infrastructure graph with a React Flow canvas
• Cypher queries — run arbitrary graph queries from the chat or graph explorer
• Multi-tenant isolation — label-based scoping ensures data separation between organizations

AI Chat with Tool Calling

A streaming conversational AI with access to every connected system:

• Graph tools — search nodes, traverse relationships, run Cypher queries
• Connector tools — execute live API calls to GitHub, AWS, K8s, databases, and more
• Tool catalog — meta-tools that let the LLM discover and invoke connector tools on demand, instead of loading all tools upfront
• Approval flow — the agent requests permission before executing sensitive operations, with configurable timeouts
• Policy enforcement — every tool invocation is checked against zero-trust policies before execution
• Streaming SSE — real-time token streaming with reasoning blocks, tool calls, and status updates

Supports multiple LLM providers:

Provider	Models	Notes
Anthropic	Claude Sonnet 4, Claude Opus 4	Recommended. Best tool-calling performance
OpenAI	GPT-4o, GPT-4.1	Full support including streaming
Ollama	Llama 3, Mistral, Qwen, etc.	Local inference, no API key required
vLLM	Any supported model	Self-hosted, OpenAI-compatible
LM Studio	Any GGUF model	Desktop local inference
llama.cpp	Any GGUF model	Lightweight local inference
Vercel AI Gateway	Multiple providers	Unified gateway

Sub-Agent Orchestration

Build complex automation with the OpenFang sidecar:

• Visual workflow editor — drag-and-drop canvas with conditional branching, loops, and parallel execution
• Sub-agent spawning — create specialized agents with isolated tool sets and contexts
• Triggers — cron schedules, webhooks, and lifecycle events
• Gateway management — route requests to different AI providers with load balancing
• Templates — pre-built workflow templates for common automation patterns

Zero-Trust Security

Enterprise-grade policy enforcement powered by OPA/Rego:

• In-process policy engine — uses regorus (Rego in Rust), no OPA sidecar needed
• Default-allow model — policies define deny[msg] rules; any denial blocks the action
• 15 granular permissions — from connectors:read to policies:manage
• Role-based access control — define custom roles mapping to permission sets
• Audit logging — every policy evaluation recorded with full context
• Policy sources — sync policies from GitHub repos or HTTP endpoints
• Policy simulator — test rules against sample inputs before deployment
• API and tool enforcement — middleware checks on every HTTP request, plus pre-execution checks in the agent loop

Distributed Workers

Run long-running jobs and remote execution across your infrastructure:

• Redis-based job queue — reliable task distribution with at-least-once delivery
• Heartbeat monitoring — track worker health and auto-reassign stalled jobs
• Remote execution — run commands on remote hosts through registered workers
• Horizontal scaling — add workers to handle increased load

Deployment Options

Agentcy runs wherever you need it:

Option	Best For	Setup Time
Docker Compose	Development, small teams	5 minutes
Desktop App (macOS)	Individual use, demos	2 minutes
Railway	Quick cloud deployment	10 minutes
Fly.io	Edge deployment	10 minutes
AWS (ECS/EKS)	Production, enterprise	30 minutes
Kubernetes / Helm	Large-scale, self-hosted	30 minutes

Feature Comparison

How Agentcy compares to alternative approaches:

Capability	Agentcy	Backstage	Port	Kubecost	Custom Scripts
Knowledge graph	Yes (Neo4j)	No	No	No	No
AI chat with tool calling	Yes (multi-LLM)	No	Limited	No	No
Live connector tools	15+ sources, 100+ tools	Plugins (read-only)	Integrations	K8s only	Manual
Visual graph explorer	Yes	No	No	No	No
Sub-agent orchestration	Yes (OpenFang)	No	No	No	No
Zero-trust policies	OPA/Rego, 15 permissions	Basic RBAC	RBAC	N/A	N/A
Semantic search (RAG)	Local embeddings	No	No	No	No
Desktop app	Tauri v2 (macOS)	No	No	No	No
Self-hosted LLM support	Ollama, vLLM, llama.cpp	N/A	No	N/A	N/A
Deployment model	Self-hosted or Agentcy Cloud (PaaS)	Self-hosted	SaaS	Self-hosted	DIY

Technology Stack

Layer	Technology
Backend	Rust, Axum, Tokio, sqlx
Frontend	Next.js 16, React 19, Tailwind CSS v4, shadcn/ui
Graph DB	Neo4j 5 (with APOC)
SQL DB	PostgreSQL 16
Cache / Queue	Redis 7
Desktop	Tauri v2 (macOS, WKWebView)
AI	Claude, GPT-4o, Ollama, vLLM, LM Studio, llama.cpp
Policies	OPA/Rego via regorus
Embeddings	fastembed (local, no API)
Orchestration	OpenFang (optional sidecar)

Who Is It For?

• Platform Engineering teams managing complex multi-cloud infrastructure
• DevOps / SRE teams needing unified visibility and incident response across tools
• Security teams enforcing access policies and auditing actions across systems
• Developers building AI-powered internal tools and automations
• Data Engineers exploring relationships across databases, APIs, and services
• Startup CTOs who want one tool instead of ten dashboards

Next Steps

• Getting Started — set up Agentcy in under 5 minutes
• Core Concepts — understand the architecture
• Docker Quick Start — detailed Docker Compose setup
• Desktop App — install the macOS desktop app