Skip to main content

Open source · OptimumAI

agentx-kit
for builders

The same framework startups promote as their core — except it's MIT, on PyPI, in VS Code, and wired into AgentX Control Plane when you outgrow a single repo.

Open source · agentx-kit

MIT · PyPI · VS Code · MCP

The open-source agentic framework behind OptimumAI AgentX.

agentx-kit is OptimumAI’s MIT-licensed, provider-agnostic framework and project scaffolder for LangChain and CrewAI. Builders pick a provider, agents, RAG, memory, MCP tools and skills — then get a ready-to-run uv project. The Control Plane is the enterprise ops layer on top; the kit is what communities and startups ship with day one.

v1.4.0 · MIT · Python ≥3.10, <3.15 · import agentx · CLI agentx

Scaffold in minutes

agentx new (or the VS Code wizard) generates a structured uv project — nodes, state, schemas, prompts, RAG utils — not a single opaque script.

12 LLM providers

Swap OpenAI, Azure, Anthropic, Gemini, Bedrock, Ollama and more without rewriting agent graphs. Env vars and model catalogs are first-class.

Real multi-agent patterns

Supervisor, sequential, or parallel orchestration; sub-agents as tools; deep agents with planning, sandbox tools, and optional critic loops.

Evals, guards & cost

evaluate_run, prompt analysis, PII guardrails, usage limits, caching and a Streamlit dashboard — the same primitives AgentX Control Plane surfaces at scale.

VS Code for builders

AgentX Kit extension

Scaffold, visualize, and run AgentX agentic projects (LangGraph / CrewAI) from VS Code. It shells out to the agentx CLI in your workspace .venv (or PATH) — no bundled Python, always the same toolkit as CI.

Marketplace ID OptimumAi.agentx-kit · v0.1.0

Install from Marketplace
  • AgentX: New Project…
    Guided pick-list → agentx new --yes in the terminal
  • AgentX: Show Graph
    Mermaid agent graph in a VS Code webview
  • AgentX: Launch Dashboard
    Opens the prompt observability UI on :8501
  • AgentX: Run Project
    uv run <slug> from agentx.json
  • AgentX: List Providers
    Providers + required env vars in the Output channel
Capabilities
RAG that ships

Chunk + embed with FAISS or Chroma, eight embedding backends (including local HuggingFace), loaders for PDF/Excel/CSV/Word, incremental re-index via manifest.

Flow as a DAG

agentx flow turns a file or whole project into a call graph — static AST, live traces, Mermaid/DOT export, or an interactive 2D/3D viewer.

MCP for IDEs & agents

agentx mcp exposes the kit to Claude, Copilot and Codex so scaffolding and project ops live inside your coding agents.

Azure-native path

agentx azure scaffolds and deploys AI pipelines to Azure ML, Container Apps, Cosmos, Service Bus and Key Vault templates.

Voice & channels

Optional STT/TTS (faster-whisper / edge-tts) and Claw multi-channel webhooks for WhatsApp, Telegram, Slack and email.

Enterprise starter pack

agentx new --enterprise adds tracing, guardrails, FastAPI, Docker, CI, evals and caching — a production skeleton startups can fork.

Providers

One catalog, twelve backends — pick in the wizard, VS Code, or agentx providers.

OpenAIAzure OpenAIOpenRouterAnthropicGeminiVertex AIBedrockGroqOllamaHuggingFaceCohereMistral
Insights for communities & builders

Why frameworks win adoption

Defaults beat tutorials

Teams don’t stall on ‘which folder structure?’ — agentx new encodes OptimumAI’s production layout (nodes/, state/, schemas/, prompts/, utils/) so every repo looks familiar to reviewers and CI.

Provider lock-in is a tax

Swap models without rewriting graphs

A curated provider catalog plus resilient tool-call parsing for small/local models means you can cost-down to Ollama or scale up to frontier APIs without a rewrite sprint.

Ops starts in the kit

Same primitives as the Control Plane

Token/cost estimates, LLM-judge evals, PII redaction and usage limits are not afterthoughts — they are Python APIs the enterprise Control Plane already consumes for BFSI and healthcare workflows.

Meet builders where they work

CLI · PyPI · VS Code · MCP

Ship from terminal, import agentx in notebooks, scaffold from the Command Palette, or let Claude/Copilot drive agentx mcp. Distribution breadth is how open frameworks become defaults.

Who it's for
Builders & startups

Use the kit as your agent runtime and scaffolder. Stay vendor-neutral while you find product-market fit — graduate to AgentX Control Plane when you need fleet evals, governance and AgentOps.

Open-source contributors

MIT license, Python 3.10+, public docs and issues. Extend providers, RAG loaders, domains and VS Code commands; templates upgrade via agentx upgrade.

Educators & bootcamps

Interactive wizard + demos that run without API keys. Teach orchestration, RAG and evals with a single CLI students already know from pip.

Platform teams

Standardize agent project layout across squads. Validate with agentx validate, visualize with graph/flow, enforce guards before Control Plane promotion.

Quickstart
# 1. Install
pip install "agentx-kit[all]"

# 2. See providers
agentx providers

# 3. Scaffold (no keys needed to generate)
agentx new --yes --name my-bot \
  --provider openai \
  --prompt "You are a support agent that answers from our docs."

# 4. Run
cd my-bot && cp .env.example .env   # add your API key
uv sync && uv run my-bot

# 5. Optional: dashboard + MCP
pip install "agentx-kit[dashboard]" && agentx dashboard
claude mcp add agentx-kit -- agentx mcp
Control Plane bridge

From kit → enterprise AgentOps

AgentX Control Plane runs the same kit APIs for evaluation, governance guards, prompt insights and usage limits. Prototype locally with pip install agentx-kit, then promote agents into Studio, Evaluator, Observability, Governance and AgentOps when reliability and auditability matter.