OpenGeni: Self-Hostable Agentic Runtime Released

What changed

Cloudgeni-ai has introduced OpenGeni, an open-source agentic runtime service designed for organizations. This service aims to provide a robust environment for AI agents to perform real-world tasks. Key features include durable and replayable agent sessions, where every action is logged in a PostgreSQL database, enabling full history reconstruction. Sessions are designed to continue working towards a defined goal until completion, pausing with a rationale, or being interrupted by human intervention. The system incorporates human-in-the-loop capabilities through tool approvals for risky actions, allowing agents to pause, request structured input, and resume specific tool calls, even after restarts. OpenGeni supports flexible deployment options, allowing agents to run in managed sandboxes (using Docker, Modal, or cloud providers) or directly on user-owned hardware, referred to as 'Connected Machines'. These machines only initiate outbound connections, enhancing security by not receiving direct credentials from Opengeni. The service also includes a feature for agent knowledge, providing a searchable library of files, retained sources, and findings with ownership and review controls. Integration into existing products is facilitated through a server-side chat handler and a session SDK with React components. The project is built using Bun, Hono, React, Vite, Temporal, PostgreSQL with pgvector, NATS, and the OpenAI Agents SDK, with a Rust agent and relay for Connected Machines. The entire codebase is licensed under Apache-2.0.

Why it matters for builders

OpenGeni provides builders with a foundational layer for deploying sophisticated AI agent applications. The emphasis on durable sessions and replayable logs addresses critical needs for reliability and auditability in production environments. The inclusion of human approval workflows and governed credentials allows for safer integration of AI agents into business processes that require oversight. Furthermore, the choice between managed sandboxes and self-hosting on personal hardware offers significant flexibility, catering to different security, cost, and control requirements. The open-source nature and permissive Apache-2.0 license encourage customization and integration into a wide range of applications.

Practical impact

Developers can begin experimenting with OpenGeni by signing up for the managed service at app.opengeni.ai, connecting a model provider, and starting their first session. For those preferring local development or evaluation, a quickstart guide is available, requiring tools like Bun, Git, Docker, and an OpenAI or Azure OpenAI key. The bun run dev command sets up the necessary services, including PostgreSQL, NATS, and Temporal, allowing users to interact with the agent via a local web interface. For integration into custom products, the OpenGeni SDK (available via @opengeni/sdk/chat) can be used to manage sessions, send messages, and handle agent replies programmatically. The project also offers React components for building chat UIs and a product integration guide for deeper embedding. Builders can also explore self-hosting options for production deployments, with reference Terraform configurations provided for Azure, AWS, and GCP.

Caveats and source limits

The provided source material details the features and setup of OpenGeni but does not include specific benchmark results or independent performance evaluations. While the latest release is noted as agent-v0.1.17, a precise release date for this version is not specified, though the project's published_at date is August 21, 2026. Pricing details for the managed service are not available in the provided excerpts; users are directed to sign up at app.opengeni.ai for more information. The source also mentions optional curated Skills with their own licensing, distinct from the main Apache-2.0 license of the core OpenGeni project.