Rivet – Open-source AI Agent dev environment

TL;DR

TL;DR: Rivet is an open-source, visual AI agent development environment built by Ironclad that lets you design, debug, and deploy LLM-powered workflows without locking into a single provider.

Source and Accuracy Notes

• Official site
• GitHub repository
• Documentation

What Is Rivet?

Rivet is an open-source AI agent development environment created by Ironclad, the company behind the well-known AI contract review platform. It was built to solve a practical problem: building real-world LLM-powered workflows requires iterating across prompts, tools, and model providers, and most developer tools in this space are either too simplistic or too opaque.

Rivet provides a visual node-based editor for designing AI agent graphs, with full debugging support, step-through execution, and support for multiple LLM providers (OpenAI, Anthropic, local models, etc.). It is designed for developers who want to ship reliable AI agents, not just prototype prompts.

Setup Workflow

Step 1: Install Rivet

npm install -g @ironclad/rivet

Or use the standalone executable for your OS. The editor runs as a desktop application with a web-based node editor.

Step 2: Create a New Project

rivet new my-agent
cd my-agent

This scaffolds a project with example graphs and a basic configuration.

Step 3: Connect an LLM Provider

Configure your API keys in rivet.config.json:

{
  "providers": {
    "openai": {
      "apiKey": "sk-..."
    }
  }
}

Rivet supports OpenAI, Anthropic, Azure OpenAI, Google AI, and local model endpoints via OpenAI-compatible APIs.

Step 4: Build Your First Graph

Open the project in the Rivet editor:

rivet editor

Drag nodes from the palette, connect them, and define your agent’s logic visually. The editor supports subgraphs, loops, and conditional branching.

Step 5: Debug and Iterate

Use the built-in debugger to step through graph execution, inspect intermediate LLM outputs, and trace tool call results in real time.

Step 6: Export and Deploy

Export your graph as a JSON configuration or use the Node.js library to embed Rivet-powered agents in your application:

npm install @ironclad/rivet

Deeper Analysis

Visual vs. Code-first: Rivet takes the visual approach seriously. Unlike debugging a chain of prompts in a notebook, you get a graph view where each node is a discrete step with explicit inputs and outputs. This makes complex branching logic far easier to reason about.

Multi-provider flexibility: One of Rivet’s strongest points is that you are not locked into a single LLM API. Switching between OpenAI and Anthropic mid-graph is a config change, not a code rewrite. For teams evaluating model performance vs. cost, this matters.

Open-source and self-hostable: Unlike many AI agent frameworks that are SaaS-only, Rivet is fully open-source (MIT license) and can be run locally. This aligns well with data privacy requirements and self-hosted deployment workflows.

Built by Ironclad: Ironclad uses Rivet internally for their own AI workflows. The tool is production-grade rather than a side project, which gives confidence in its stability and long-term maintenance.

Comparison to alternatives: Compared to tools like LangFlow, Rivet is more developer-focused with first-class debugging. Compared to building directly with the OpenAI Assistants API or LangChain, Rivet abstracts away boilerplate while still exposing the underlying graph structure.

Practical Evaluation Checklist

• Open-source with MIT license
• Visual graph editor with step-through debugger
• Multi-LLM provider support (OpenAI, Anthropic, Azure, Google AI, local)
• Export as embeddable Node.js library or standalone JSON config
• Desktop editor (Electron) with cross-platform support
• Active development by Ironclad engineering team
• Subgraph support for modular, reusable components

Security Notes

• API keys are stored locally in rivet.config.json, not sent to any external service beyond the LLM providers you configure.
• No telemetry or data collection is documented in the open-source codebase.
• For enterprise use, run Rivet in a network-isolated environment and manage secrets via your existing secrets manager.

FAQ

Q: Do I need an Ironclad account to use Rivet? A: No. Rivet is a standalone open-source tool. You only need API keys for whichever LLM providers you want to use.

Q: Can Rivet be used in a production backend? A: Yes. You can export your agent graphs as JSON and embed them in any Node.js application using the @ironclad/rivet package. It is designed for production use, not just prototyping.

Q: How does Rivet compare to LangChain or LlamaIndex? A: Rivet is a visual workflow orchestration layer, whereas LangChain and LlamaIndex are code-first libraries. If you want to prototype quickly in code, those are options. If you want a visual debugger and a graph-based mental model, Rivet fills that gap.

Q: Is there a hosted or SaaS version of Rivet? A: Not at launch. Rivet is self-hosted only. Ironclad may offer a hosted version in the future.

Conclusion

Rivet is a serious, open-source answer to the question of how to build and debug AI agent workflows without getting locked into a single provider or a single paradigm. The visual graph editor makes complex branching logic tangible, and the integrated debugger removes a major pain point in LLM development. If you are building AI agents in production and want a tool that grows with you, Rivet is worth evaluating.