Tool Pick｜Opengeni — session persistence, human approval, and credential governance for agents, bundled into one self-hostable service

🌏 中文版

Tool Info

What problem it solves

Build an agent product that needs to keep running for a while — a support agent that remembers state across conversations, an infra agent that needs a human to approve a change before it ships, a research agent that runs for hours and can't afford to restart from scratch — and you quickly find the model itself only covers a fraction of what's needed. A model is a function from tokens to tokens: it forgets everything between calls, has no idea what it's allowed to do, and has no obligation to keep working until a job is actually done. Where sessions get stored, how a browser refresh reconnects to the same conversation, how a risky tool call gets gated behind a human's approval, how the credentials an agent can see stay separate from the prompt it's executing — nearly every team building an agent product ends up stitching this together themselves, usually out of Postgres, some workflow engine, and a hand-rolled approval UI.

Opengeni's approach is to package that surrounding layer into a ready-made, self-hostable service: every event lands in Postgres as a replayable record, so a browser reload or a brand-new client can pick up the same history. A session is given a goal with success criteria, and the agent keeps working until it meets that goal, pauses with a rationale, or a human interrupts it. Tool calls can sit behind an approval gate, and an agent can even resume the exact interrupted tool call once a human answers its question. Execution can run in a managed sandbox, or on a machine you register yourself (a Connected Machine) — the key detail being that machine only ever dials out and never holds any Opengeni credentials. The project's own framing is "rent the edges, own the middle": models, provider APIs, and raw compute change too fast to be worth owning, but session state, governance rules, and the knowledge base are the part that's actually yours and belongs in a Postgres database you control.

Good fit: teams already building agent products that need to run long, need a human in the loop, and need an audit trail, and who'd rather use ready-made infrastructure than assemble their own out of Temporal and Postgres; or teams who want to validate product shape first on the zero-deployment managed version before deciding whether to self-host. If you're only writing a small, single-shot agent that finishes in a few seconds, this whole layer of persistence and approval gating is more than you need — a plain agent-logic framework like LangGraph or CrewAI is lighter weight.

Quick Start

Install

# Local dev/evaluation: needs the exact Bun version in .bun-version, Docker, rustup, a C compiler
git clone https://github.com/Cloudgeni-ai/opengeni.git
cd opengeni
cp .env.example .env   # add your OpenAI or Azure OpenAI key
bun run dev

bun run dev installs dependencies, starts Postgres, NATS, Temporal, and object storage, runs migrations, and starts the API, workers, and web app; open http://127.0.0.1:3000, describe a task, and watch the session run. If you'd rather skip the dependency stack, sign up directly at app.opengeni.ai, connect a model, and start a session with nothing to deploy.

Basic usage

Wiring Opengeni into your own product takes one server route and one component:

// app/api/opengeni/[...path]/route.ts (server only; the key stays here)
import { Opengeni } from "@opengeni/sdk/chat";
import { createSessionProxyRoute } from "@opengeni/sdk/next";

const og = new Opengeni({ apiKey: process.env.OPENGENI_API_KEY! });

export const { GET, POST, PUT, PATCH, DELETE } = createSessionProxyRoute(og, {
  resolve: async (request) => {
    const me = await authenticate(request); // your product's own auth
    if (!me) return new Response("Unauthorized", { status: 401 });
    return { user: me.id, tenant: me.teamId };
  },
  createSession: (input) => input,
});

// Browser
import { OpenGeniChat } from "@opengeni/react";
import "@opengeni/react/compiled.css";

Advanced usage

Let a coding agent like Claude Code wire Opengeni into an existing project for you:

claude plugin marketplace add Cloudgeni-ai/opengeni
claude plugin install opengeni@opengeni --scope user

Once installed, the agent connects to your organization over MCP (sign in once in the browser) and gains the bundled Opengeni skills that handle the integration details.

Compared to existing approaches

Things to watch out for

• Self-hosting isn't lightweight: local development alone needs a specific Bun version, Docker, rustup, and a C compiler; running it in production means standing up Postgres, NATS, and Temporal — this isn't a single-binary, one-line install. If you just want to feel out the product shape, the zero-deployment managed version is faster.
• Follow the README's security boundary before going to production: it's explicit that you should "not expose a production deployment without a deliberate access mode, tested database role posture, rate limits, and a reviewed sandbox credential policy" — details live in the security boundary section of docs/deployment.md.
• Stars and project maturity are still early: 198 stars, 49 open issues, created in April 2026 — for the scope it covers (identity, tenancy, permissions, approvals, audit, credential governance all at once), this is an ambitious, still fast-moving project. Watch its issues and release history for a while before depending on it in production.

Today's takeaway

Most discussions of "how to build an agent product" still center on picking an agent framework or a prompt strategy, but what actually gets an agent into a real product is usually the layer outside the framework: where sessions live, when a human needs to step in, how credentials stay out of the agent's own view. Opengeni calls that layer out specifically and packages it as something you can just use — turning "infrastructure every agent product has to reinvent" from each team's individual tech debt into a shared layer you can choose whether to maintain yourself.

References

• Cloudgeni-ai/opengeni — GitHub
• Opengeni website
• Opengeni self-host docs (security boundary, deployment guide)
• Opengeni official blog
• Model Context Protocol specification