ADEPT¶

Agentic Discovery and Exploration Platform for Tools

ADEPT is an open-source platform for building secure, production-ready AI agent systems with integrated scientific tooling. Built on the Model Context Protocol (MCP) and LangGraph, it provides a complete framework for deploying LLM-powered agents across multi-cloud infrastructure.

Key Capabilities¶

• 28+ Built-in Scientific Tools -- BLAST, UniProt, PubChem, AlphaFold, web search, sandboxed code execution, and ChromaDB-backed RAG pipelines
• Multi-Agent Orchestration -- Router mode (supervisor dispatches to specialized workers) and Graph mode (LangGraph DAGs with dynamic state transitions)
• LLM-Agnostic -- Seamless provider switching via LiteLLM: Azure OpenAI, AWS Bedrock, Anthropic, Ollama, NVIDIA NIMs, and more
• 10 Client Interfaces -- Claude Desktop, Claude Code, ChatGPT, Roo Code, Continue.dev, Cursor, OBot.ai, AG-UI, OpenWebUI, and JupyterLab through a single API surface
• A2A Federation -- Agent-to-Agent protocol for cross-organization agent sharing and gateway mesh communication
• Enterprise Security -- Keycloak OIDC/JWT authentication with per-tool ACL-based RBAC and sandboxed code execution via nsjail
• Multi-Cloud Deployment -- AWS CDK, Azure Pulumi, GCP Terraform; Helm charts for Kubernetes; Docker Compose for local development
• Observability -- Full request tracing and token cost tracking via Langfuse integration

Architecture¶

ADEPT implements a three-tier secure architecture separating authentication, orchestration logic, and tool execution:

graph TD
    subgraph Clients
        A[Claude Desktop / Code]
        B[OpenWebUI / JupyterLab]
        C[ChatGPT / Cursor / Roo Code]
    end

    subgraph "Tier 1: Agent Gateway"
        GW[Authentication Proxy<br/>Keycloak JWT + RBAC]
    end

    subgraph "Tier 2: Orchestration Service"
        OS[LangGraph Agent Brain<br/>PostgreSQL State + Multi-Agent]
    end

    subgraph "Tier 3: MCP Tool Servers"
        MCP1[Scientific Tools<br/>BLAST, UniProt, PubChem, RAG]
        MCP2[HPC Pipelines<br/>Nextflow, Video, GitXray]
        MCP3[Sandbox Execution<br/>Secure Code via nsjail]
    end

    subgraph Infrastructure
        KC[Keycloak IAM]
        PG[(PostgreSQL)]
        RD[(Redis)]
        LF[Langfuse Tracing]
    end

    A --> GW
    B --> GW
    C --> GW
    GW --> OS
    OS --> MCP1
    OS --> MCP2
    OS --> MCP3
    OS --> PG
    OS --> RD
    GW --> KC
    OS --> LF

:octicons-arrow-right-24: Full Architecture Overview

Quick Start¶

# 1. Clone the repository
git clone https://github.com/pnnl/adept-agentic.git
cd adept-agentic

# 2. Configure environment
cp .env.example .env
# Edit .env with your LLM provider credentials (Azure OpenAI, AWS Bedrock, etc.)

# 3. Start the platform
make start

Once running, the Agent Gateway is available at http://localhost:8083/v1 with an OpenAI-compatible API.

pip install adept-connectors

:octicons-arrow-right-24: Getting Started Guides

Client Interfaces¶

ADEPT exposes an OpenAI-compatible API surface, enabling integration with a broad set of AI clients:

:octicons-arrow-right-24: Connectors Guide

Documentation Sections¶

Technical Report¶

Rigor, P. et al. "ADEPT: A Pedagogical Framework for Integrating Agentic AI with Deterministic Scientific Workflows." Pacific Northwest National Laboratory, 2026.

:octicons-arrow-right-24: Read on OSTI.gov

License¶

BSD 2-Clause License. Copyright Battelle Memorial Institute 2026.

:octicons-arrow-right-24: View License