sarkar4777/abenix

The open-source AI agent platform for problems chatbots can't solve.
Graph-grounded knowledge · governed decisions · production-grade orchestration · cloud or edge.
Why Abenix  · 
Architecture  · 
Quick start  · 
Showcase apps  · 
Deploy  · 
Enterprise readiness
---
> **Run it now** on a laptop with Docker, minikube, kubectl and Helm. One command builds and starts everything, see Quick start for what you need first.
>
>
🧭 Contents
- How it fits together
- Why Abenix
- Architecture
- Quick start
- What you need first
- One command on minikube
- Other ways to run it
- Where everything lives
- Required env vars
- Showcase apps
- OracleNet — strategic decision-analysis
- Mideast Tourism — Vision-2030 analytics
- ClaimsIQ — insurance claim adjudication (Java)
- Industrial-IoT — predictive maintenance + cold chain + edge
- ResolveAI — customer-resolution case management
- ContractIQ — energy contract intelligence
- PharmaVigil — drug-safety intelligence
- Wingman — commodities trading desk
- Edge runtimes
- How a cloud agent reaches the edge
- Pick a variant
- Signing keys, fail closed
- Lifecycle
- Deploy anywhere
- Build on top of Abenix
- Enterprise readiness
- Tech stack
- Documentation
- Contributing
- License
---
The story
Most AI agent platforms give your agents amnesia — they retrieve documents, forget context, and re-derive the world model on every turn. Most also assume "agent" means "a chatbot in the cloud talking to OpenAI." That's fine for support tickets. It's not fine for a wind turbine, a refrigerated trailer, a contract worth seven figures, or a control room with a 50 ms hard limit.
**Abenix gives agents a brain, a body, and a rulebook.** A typed graph that lives next to the knowledge base so agents traverse evidence like a researcher follows citations. A deterministic rule engine that makes the decisions an agent should not make on its own, with risk tiers that decide who signs off and what an agent may touch. Per-agent runtime pools, multi-signoff approvals, idempotency, and a dead-letter queue so you actually trust them in production. And lean edge runtimes — Python, Rust, or C — that take the same agent definition, sign it into a 12–80 MB bundle, and run it next to the equipment.
Same agent. Same definition. Cloud or edge. Built for the long-running, knowledge-heavy, accountability-mandatory work that workflow tools choke on.
A fresh install from the one-command setup, signed in for the first time with a model connected: the Start here checklist and live activity
---
🗺 How it fits together
New here? Read How Abenix fits together first. It explains the core objects (agents, pipelines, tools, knowledge, decisions, approvals, autonomy, improvements, runs), what each role does, every word the UI uses, and where to start.
In the app, three things keep you oriented:
- **Start here** on Home is a short checklist for your role. Admins connect a model and invite the team, creators build, test and improve an agent, members chat and give feedback.
- **Needs you** counts everything waiting on you: approvals, proposed fixes, watching reviews, held content, marketplace submissions and new alerts.
- The sidebar starts in **Essentials**, a short list for daily work. **Show all tools** opens every page, grouped as Build, Run & Test, Monitor, Marketplace, Admin and Workspace. Every page opens with a one-line purpose, its main action and a "How this works" panel, and the **Help** page covers every screen.
---
✨ Why Abenix
Eleven things that, taken together, you do not get anywhere else open-source:
1. Graph-grounded knowledge — Atlas + Knowledge Engine
Documents and concepts live on the same canvas. Drop a PDF → multimodal extraction proposes typed nodes and edges with confidence scores. Type a sentence → cardinality inference. Time-slider snapshots the whole graph on every save. Five starter ontologies in the box (FIBO Core, FIX Protocol, EMIR, ISDA, ETRM EOD).
Agents query the graph through five typed tools — atlas_describe, atlas_query, atlas_traverse, atlas_search_grounded and atlas_as_of — and get back **paths of cited evidence**, not three similar paragraphs. The vanilla-RAG comparison:
| Question | Vanilla RAG | Abenix |
|---|---|---|
| *"What caused the Q3 revenue drop?"* | 3 similar paragraphs | Q3 Report → mentions → supply chain delays → CAUSED_BY → chip shortage |
| *"Counterparties with > 5 unconfirmed trades in 7 days"* | Cosine miss | Pattern walk over the typed graph, structured rows back |