Opengeni
Cloudgeni · Opengeni: Open-Source, Self-Hostable Runtime for Long-Running AI Agents in Your Product
Opengeni is an open-source (Apache-2.0) runtime and control plane for long-running AI agents. It keeps each session in a replayable Postgres event log, pauses for human approval before risky tool calls, keeps credentials out of the prompt, and runs work in a managed sandbox or on a machine you enroll. It suits product and platform teams that want to embed agents in their own app, or give them to their team, without building that infrastructure themselves.
Best for
Product and platform teams building agent features into their own TypeScript/React product, or running agents for their organization, who want durable sessions, approvals, credential handling, and the choice to self-host under Apache-2.0 or start on a hosted cloud with the same API
Not ideal for
Teams that want a ready-made agent rather than infrastructure, Python-only backends that want a native SDK, small teams unwilling to operate Postgres, Temporal, NATS, and Kubernetes if they self-host, and buyers who need a mature, slow-moving platform today
Who it's for
Product engineers embedding agents in a SaaS app, platform teams running a shared agent runtime for their organization, and companies that want agent session history, approvals, and audit data in their own database
Capabilities
- Durable sessions: every event is written to a Postgres log with a sequence number, live streams backfill from it over Server-Sent Events, and a crashed worker's turn is retried as a new attempt without repeating the prompt
- Goals with success criteria keep a session working across turns until it completes with evidence, pauses with a reason, or a person interrupts
- Tool approvals with Allow, Ask first, and Block choices per connected account; approving resumes the exact stored call, and only humans can approve
- Built-in `request_human_input` tool for up to 20 free-text, single-select, or multi-select questions, with optional Skip and deadline
- Send, Steer, Pause, Resume, and Cancel controls, with queued messages that stay editable until they start
- Compute per session: a managed sandbox (Docker, Modal, or a cloud sandbox provider), no compute, or a Connected Machine you enroll, which connects out with no inbound ports and uses its own Git and SSH credentials
- Stock sandbox images include Terraform, Checkov, GitHub CLI, git, psql, a Python toolchain, and a Chromium-based browser, and session files can be recovered from a snapshot after the sandbox is gone
- Connected accounts scoped to Only me or This workspace, plus a credential provider endpoint that returns short-lived environment variables, files, Git, or MCP credentials before each run
- Product tools via your own MCP server or an OpenAPI description compiled into tools, called as the signed-in user with short-lived tokens the session proxy mints
- Knowledge library of files, sources, findings, and collections, with workspace instructions and Skills, and Automatic, Review first, or Off settings for what agents may change
- Embedding kit: `@opengeni/sdk` (sessions, workspaces, tools, files, event streams), a packaged session proxy for Next.js, Express, and Hono, `@opengeni/react` components themed with CSS tokens, and a chat facade for existing Vercel AI SDK or OpenAI-shaped chat UIs
- Background agents from schedules or signed webhook events, with outbound webhooks when a turn completes, fails, or waits for input
- Workspaces mapped to your product's tenants and users, usage allowances per team and person, and Insights on spend, model calls, and tokens
- Plugin for Claude Code, Codex, and Cursor that gives your coding agent Opengeni integration skills and MCP access to your organization
- Self-host with a Helm chart published to GHCR and reference Terraform for AWS, Azure, and GCP
Limitations
- A young project: the GitHub repository was created in April 2026, the public launch was October 5, 2026, and release artifacts are published several times a day (server release candidates were at 0.23.x in early October 2026)
- Self-hosting is a substantial stack: the API, web app, two Temporal worker types, Postgres with pgvector, Temporal, NATS, object storage (Azure Blob, S3, or GCS), and a sandbox backend. The Helm chart's built-in copies of those services are meant only for local and smoke testing
- The deployment guide warns that some migrations require stopping every old API and worker first, and that the old image must not be restarted afterward
- The README says the local development stack (`bun run dev`) runs agent commands directly on your machine by default (the `local` sandbox) unless you set the Docker sandbox, and it needs Bun, Docker, rustup, a C compiler, and an OpenAI or Azure OpenAI key
- The SDK and React components are TypeScript only. Other backends (Django, Rails, Go, PHP, Java) proxy the HTTP API by following the docs, and no Python SDK is documented
- Connected Machines are off by default on self-hosted deployments and need a stream relay and NATS auth-callout to be set up
- The README warns not to expose a production deployment without a deliberate access mode, tested database role posture, rate limits, and a reviewed sandbox credential policy, and the self-hosting guide notes that the shared deployment key is a simple perimeter rather than per-tenant access control
- Sessions run Opengeni's own agent loop (built on the OpenAI Agents SDK) with the models you connect. The public docs do not describe running existing harnesses such as Claude Code or Codex as the session agent
- On Opengeni Cloud, a turn ends when prepaid credits run out. Usage through your own connected key or subscription is billed by that provider rather than drawn from Opengeni credits, but the pricing page does not say whether any Opengeni fee applies to it
Use cases
- Adding a billing or support assistant to a SaaS app that looks up invoices through the product's own MCP tools and asks the user before issuing a refund
- Running a nightly reconciliation agent from a schedule with a goal and success criteria, and getting a webhook when it finishes
- Pointing a session at an enrolled build server or GPU box so the agent works with the files, Git credentials, and tools already there
- Giving an internal team a shared agent workspace where sessions, approvals, and learned Knowledge are visible and reviewable
- Starting agent sessions from a product's events through signed webhooks, with each run receiving short-lived cloud credentials from the product's backend
Our take
Opengeni is a serious attempt at the unglamorous layer that turns a model into a dependable product feature: a replayable event log, approvals that resume the exact call, goals that keep work going, and credentials that stay out of the prompt. Connected Machines treat hardware you own as a first-class compute target alongside managed sandboxes, and because the hosted cloud runs the same Apache-2.0 code and deployment tooling you can self-host, starting hosted does not lock you in. The costs are maturity and operations. It only launched publicly in October 2026, release artifacts arrive several times a day, self-hosting means running Temporal, NATS, and Postgres, typically on Kubernetes, and the SDK is TypeScript-only. Teams with a TypeScript product and some platform capacity should try the hosted app or the local stack on one real workflow before committing.
Who should use it
TypeScript and React product teams that want agents inside their app with approvals and user-scoped tools, platform teams that want a self-hosted agent runtime whose session history and audit trail live in their own Postgres, and organizations that want agents working on their own build servers or GPU machines.
Who should skip it
Teams that need a finished agent rather than a runtime, Python-first shops expecting a native SDK, teams that want to run Claude Code or Codex as-is in hosted sandboxes, and organizations that need a long track record and slow, predictable upgrades.
Strengths
- Apache-2.0 down to the Helm chart and Terraform, and the same code runs the hosted cloud, so you can start hosted and move to your own infrastructure
- Replayable Postgres event log gives every client and audit the same session history and lets turns survive worker restarts
- Approvals, structured questions, and goals are built in rather than left to your application
- Connected Machines let agents work on hardware you own without opening inbound ports
- Hosted pricing is simple: model cost plus 5%, with no seat or platform fee and $10 in trial credits for new verified accounts
Weaknesses
- Very new, with fast release cadence and upgrades that can require full stops
- Self-hosting requires operating Postgres, Temporal, NATS, object storage, and a sandbox backend
- TypeScript-only SDK and React-first UI components
- Runs its own agent loop rather than hosting existing coding-agent harnesses
Opengeni pricing
Open source
$0
- Full Apache-2.0 source, including Helm chart and Terraform
- Self-host on your own infrastructure
- Bring your own model credentials
- Community support
Opengeni Cloud
Model cost + 5%
- Managed infrastructure and updates
- Supported models billed through Opengeni
- Sandboxed agent execution
- Audit and usage controls
- No seat fees, platform fee, or minimum commitment
Enterprise
Custom
- Private infrastructure requirements
- Deployment and integration help
- Security, access, and SSO requirements
- Commercial support options
Free tier limits: Self-hosting is free apart from your own infrastructure and model costs. New verified Opengeni Cloud accounts get $10 in trial credits.
Note: Opengeni Cloud adds 5% to the provider's model cost, so $100 of provider usage costs $105. Usage is paid from prepaid credits, and when credits run out the turn ends until you top up. When you connect your own model key or subscription, the docs say usage is billed by that provider and does not draw on Opengeni credits; the pricing page does not state whether any Opengeni fee applies in that case.
Technical specs
API pricing
Opengeni Cloud: provider model cost plus 5%, paid from prepaid credits, with no seat or platform fee. Self-hosted: no license fee
Available models
Where Opengeni excels
An in-product billing or support assistant
The React conversation component and session proxy put the agent in your UI, your MCP tools act as the signed-in user, and Ask first approvals stop actions such as refunds until the user confirms.
Long-running background work that has to finish
A scheduled or webhook-triggered session with a goal keeps working across turns, survives worker restarts through the event log, and notifies your backend by webhook when it completes or needs input.
Agents on hardware you already own
Connected Machines let a session run on an enrolled laptop, build server, or GPU box with its existing files and Git credentials, connecting out so no inbound ports are opened.
Opengeni vs. competitors
Opengeni vs. TrueForge
TrueForge is TrueFoundry's MIT-licensed agent harness with persistent sessions, MCP tools, approvals, schedules, and a chat UI, API, and SDK. It documents Daytona as its sandbox provider and runs locally on SQLite or hosted on Postgres and Redis, with a custom-priced managed version. Opengeni covers similar ground with a heavier Temporal, NATS, and Postgres stack, and also offers Connected Machines, goals with success criteria, a reviewable Knowledge library, and a hosted cloud with published pricing (model cost plus 5%) rather than custom quotes.
Opengeni vs. DigitalOcean Managed Agents
DigitalOcean Managed Agents is a hosted service that runs existing harnesses such as Claude Code, Codex, OpenCode, Hermes, or LangGraph apps in Firecracker microVMs, with a 16,000-tool MCP gateway and usage-based compute billing, but no self-hosted option. Opengeni runs its own agent loop, is open source and self-hostable, and focuses on embedding sessions in your product with React components and user-scoped tools. DigitalOcean fits teams that want to run an agent they already use, and Opengeni fits teams building agent features they control end to end.
Frequently asked questions
What is Opengeni?
Opengeni is an open-source runtime for long-running AI agents. It provides the layer around the model: durable, replayable sessions, goals, human approvals and questions, tools through MCP or OpenAPI, credential handling, a Knowledge library, and compute in a managed sandbox or on a machine you enroll. You can use it from its web app or embed it in your own product through its API, TypeScript SDK, and React components.
Is Opengeni open source?
Yes. Opengeni is licensed under Apache-2.0, including the API, web app, workers, Helm chart, and reference Terraform. Some optional curated Skills in the repository carry their own license metadata.
How much does Opengeni cost?
Self-hosting has no license fee; you pay for your own infrastructure and model providers. Opengeni Cloud charges the provider's model cost plus 5%, with no seat fee, platform fee, or minimum commitment, and new verified accounts get $10 in trial credits. An Enterprise option for private infrastructure, SSO, and commercial support is priced by conversation with the team.
What do I need to self-host Opengeni?
A production deployment runs the API, web app, and two Temporal workers, backed by Postgres with pgvector, Temporal, NATS, object storage (Azure Blob, S3, or GCS), and a sandbox backend such as Docker, Modal, a cloud sandbox provider, or Connected Machines. Opengeni publishes a Helm chart and reference Terraform for AWS, Azure, and GCP, and recommends managed services for the stateful pieces. For local evaluation, one command (`bun run dev`) starts the whole stack.
Which models does Opengeni support?
OpenAI and Azure OpenAI are built in, and other OpenAI-compatible endpoints can be added by configuration. The docs also cover OpenRouter, Vercel AI Gateway, Anthropic API keys, and connecting ChatGPT/Codex or SuperGrok subscriptions. Admins choose which models each workspace can use, and a session can switch models mid-conversation.
What is the difference between a managed sandbox and a Connected Machine?
A managed sandbox is created per session by Opengeni, clones repositories into `/workspace`, and runs inside the deployment. A Connected Machine is a computer you own and enroll once; the agent works in a folder on it with that machine's own Git and SSH setup, and the machine connects out to Opengeni, so it needs no inbound ports. Screen control is a separate consent, and you can revoke a machine from the Machines page.
How do I add Opengeni to my product?
Add one server route that proxies the session API with your organization key, mapping your users and tenants to Opengeni workspaces, and render `OpenGeniChat` from `@opengeni/react` in the browser. The agent reaches your product's data through an MCP server or OpenAPI description, called as the signed-in user. A plugin for Claude Code, Codex, and Cursor can do the wiring for you.
Opengeni vs TrueForge: what is the difference?
Both are open-source, model-neutral runtimes with persistent sessions, MCP tools, sandboxes, approvals, and embeddable UI. TrueForge is MIT-licensed, documents Daytona as its sandbox provider, and has a lighter self-hosted footprint (SQLite locally, Postgres and Redis hosted). Opengeni is Apache-2.0, runs on a Postgres event log with Temporal orchestration, also offers Connected Machines, goals with success criteria, and a reviewable Knowledge library, and has a hosted cloud with published pricing (model cost plus 5%), at the cost of a heavier stack to self-host.
How is Opengeni related to Cloudgeni?
Opengeni is built by the team behind Cloudgeni, an agentic CloudOps platform, and is published under Cloudgeni's GitHub organization. The README says Opengeni grew out of two years of running agents against production cloud infrastructure at Cloudgeni.
Is Opengeni ready for production use?
The README describes it as production-ready, but the project is young: the repository was created in April 2026, the public launch was October 5, 2026, and releases ship several times a day. Before exposing a self-hosted deployment, the docs ask for a deliberate access mode, gateway TLS and authentication, rate limits, and a reviewed sandbox credential policy, and some upgrades require stopping every old API and worker first. Pilot it on one real workflow before relying on it.
Integrations & fit
Alternatives to consider
TrueFoundry
Platform and application teams that want a model-neutral, self-hosted runtime for reusable production agents
Free
DigitalOcean
Developers and small teams who already use Claude Code, Codex, OpenCode, Hermes, or LangGraph and want those agents running durably in the cloud, with pause, resume, and fork, governed tool access, and usage-based billing in one place
PaidAbout Opengeni
Opengeni describes itself as not the agent but everything the agent needs around it. You give it work from its web app, or your product calls the same session API through the TypeScript SDK, the React components (`OpenGeniChat`, `SessionConversation`), or plain HTTP, with your backend holding the API key and proxying the session routes. Every session event is stored in Postgres with a sequence number, so a reload, a second client, or an audit replays the same history, and a turn whose worker dies is retried as a new attempt rather than a repeated prompt. Sessions can be given a goal with success criteria and keep working across turns until they finish with evidence or pause with a reason. Tool calls can be set to Allow, Ask first, or Block per connected account, and agents can ask structured questions; both waits survive restarts and resume the exact turn. Each session runs in a managed sandbox (Docker, Modal, or a cloud sandbox provider), with no compute at all, or on a Connected Machine: a laptop, build server, or GPU box you enroll that connects out and uses its own Git and SSH setup. Your product's data reaches the agent through an MCP server or an OpenAPI description, authorized as the signed-in user with short-lived tokens, and a credential provider endpoint can hand each run short-lived secrets that reach the sandbox but not the conversation or logs. A Knowledge library stores findings, sources, instructions, and Skills, with optional review before agent changes are published. Models are chosen per workspace and per session: OpenAI and Azure OpenAI are built in, and the docs also cover OpenRouter, Vercel AI Gateway, Anthropic API keys, ChatGPT/Codex and SuperGrok subscriptions, and other OpenAI-compatible endpoints. The project grew out of two years of running agents against production cloud infrastructure at Cloudgeni, the team's agentic CloudOps platform. Self-hosting is free; the hosted Opengeni Cloud charges the provider's model cost plus 5%, with no seat or platform fee. The tradeoffs: self-hosting means operating Postgres with pgvector, Temporal, NATS, object storage, and a sandbox backend, usually on Kubernetes; the SDK is TypeScript-only; and the project only launched publicly on October 5, 2026, with very frequent releases and upgrade steps that sometimes require stopping every old API and worker first.
Updates from Opengeni
Opengeni launched publicly on October 5, 2026 with a Product Hunt launch, where it was listed third on that day's leaderboard. The launch post positions it as an open, Apache-2.0 alternative to the managed agent platforms from cloud providers and model labs, with self-hosting free and the hosted cloud priced at model cost plus 5%.
Are you the founder? Claim this listing →