Which Managed Platforms Can Deploy Our Docker Agent, Call Models on the Same Network, and Sell It Through a Marketplace?

Listing on the Agentbox Marketplace is the last of three steps for a Docker agent: register the image from your own registry, test the live endpoint, then publish. GMI Agentbox runs the agent privately under Option 03, Compute + Models, where it calls GMI models on the same network as its container, and compute and model tokens come on one invoice. Agentbox is in early access.

The path from image to storefront

Agentbox is organized as "Deploy first. Launch when ready." Each of its four steps maps to a task in Console.

Step 1: Deploy privately from your Docker image

Choose GMI CE Deployment in the register wizard, the path where GMI hosts the container. You keep the image in your own registry, Docker Hub, GitHub Container Registry, or a private one, and GMI stores its URL and pulls from your registry each time an instance is provisioned.

The image serves HTTP on port 8080 by default, passes health checks, and stays inside the gateway timeout. Wizard screens, registry credentials, port mapping, and secrets are covered in the register guide. The wizard's container tier provides 2 vCPU, 4 GB RAM, 10 GiB ephemeral storage, and 30 GiB data storage, and that configuration pre-fills the deployment when a buyer deploys a copy. Registering stores a template without starting an instance. Until you publish, the agent stays in Draft and only you can deploy it.

Step 2: Call models on the same network

Model access is what Option 03 adds to plain hosting. The switch sits in the Infrastructure step: enable MaaS integration, then tick each model the agent is allowed to call. You can change that list later. When the container starts, GMI supplies a MaaS API key and an OpenAI-compatible base URL through its environment, which keeps model credentials out of the image. GMI_MAAS_BASE_URL, GMI_MAAS_API_KEY, and GMI_MODELS are the three variables your code reads, exactly as a standard OpenAI-compatible client would.

Because compute and inference run on the same network, the agent's model calls stay on GMI's network instead of crossing to a separate vendor. That avoids the latency, egress, and operational cost of stitching vendors together. The wizard offers GMI's 200+ models, and your selection is saved with the template, so it pre-fills when a buyer deploys a copy.

If your agent already has a model provider, you can keep it: leave MaaS integration off, store the provider key as a SECRET variable, and the agent is listed with the Powered by GMI Infrastructure badge instead.

Step 3: Validate privately before you list

Launch an instance of your own agent from the Monitor tab, or through the Integration tab's lifecycle calls, and send it sample requests. Check latency, output quality, and error handling. Each launched instance gets its own public URL. You can iterate by re-registering, and URLs stay stable across updates.

For tasks that run longer than the gateway timeout, the long-running requests guide describes the pattern of returning a job ID immediately and letting callers poll for the result.

Step 4: Create the listing

Open My Deployments & Listings, select the registered agent, and click List an agent. Starting from a registered template pre-fills the form, and a GMI CE Deployment template is detected and linked automatically. The list guide covers the form in two steps, Listing Info and Review & Publish.

Write the listing for the person who will deploy the agent. Agentbox guides buyers to evaluate a listing by sample output, documentation, publisher identity, and infrastructure badge, in that order, so put sample output and documentation up front.

When you click Publish, the listing goes live on Agentbox within minutes, and you stay responsible for keeping the description accurate. To take it down, use Unpublish from the My Agents dashboard. The agent leaves the Marketplace and existing instances keep running.

What a buyer receives

When a user deploys your agent from the listing, Agentbox provisions a copy of your registered template for them, and each end user of a deployed agent gets an isolated instance with state scoped to that session. What travels with that copy:

• Copied: container image and version, environment variables and secrets, port mappings and network config, and MaaS model selections.
• Not copied: your GMI MaaS API key (GMI injects a key at runtime), usage history or session data, and your internal project name.

Consumers reach the agent through GMI's infrastructure, so your internal endpoint is never exposed.

Badges and the Verified review

The badge on a listing tells buyers who runs the infrastructure.

Verified needs all four criteria: Identity, Infrastructure, Behavior, and Safety. Infrastructure means the agent runs entirely on GMI MaaS and GMI infrastructure; the full definitions are on the Verified agents page linked below. The badge is granted automatically when they are met and revoked automatically if any stops being met. You do not reapply after fixing an issue; it is re-granted once the criteria are satisfied again. Verified agents also get GMI-run availability, patching, and failover across regions.

Option 03 covers the Infrastructure criterion, because both layers, compute and inference, run on GMI. Identity, Behavior, and Safety still have to be met.

Who pays what once buyers deploy

Agentbox lists a commercialization path next to the Marketplace, and runtime billing is usage-based. Billing starts only when a user provisions an instance of your agent and it reaches Running status.

Registration is free, and there is no listing fee and no minimum spend. For an agent hosted on GMI, usage bills to the user's GMI account. If your agent sends users to its own product site instead, billing is handled by you as the agent provider.

Step 5: Operate after launch

The My Agents dashboard is where you run the agent once it is live. Its Monitor tab lists instance status counts and a table of instances with detail and log actions. The Integration tab gives you the template ID and the four-call lifecycle for provisioning, listing, polling, and terminating instances with your GMI API token. The Analytics tab charts usage by model over 1, 7, 30, or 90 days and links to billing. The top of the panel carries Unpublish, Edit listing, and View public listing.

FAQ

Can we list an agent that we host on our own servers and that only uses GMI for model calls?

Register it on the Self-hosted + MaaS path, where you give basics, a MaaS key, and your public endpoint, then review and submit. In the list form, enter that public URL in the Agent Access URL field, which appears on the blank form for self-hosted agents. The listing carries the Powered by GMI MaaS badge. Verified's Infrastructure criterion asks that the agent runs entirely on GMI MaaS and GMI infrastructure, which GMI CE Deployment with MaaS integration on satisfies.

Can an open-source agent be listed next to commercial agents, and can it earn the Verified badge?

Verified rests on four review points: a verified publisher account, an agent that runs entirely on GMI MaaS and GMI infrastructure, a match to its listed description and category, and no abusive prompts, data leaks, or policy violations. Open-source status is not among them. NemoClaw, an open-ecosystem agent, received a verified Agentbox listing on its first day of early access, listed alongside commercial agents rather than in a separate open-source tier.

Package, register, launch one test instance

Package the agent as a Docker image, register it with GMI CE Deployment, and turn on MaaS integration so the model calls run on GMI's network. Launch one instance and test it before you open the list form. The Agentbox overview summarizes the whole flow, and the Verified agents page has the review criteria. Start in the GMI console, or bring your listing and launch plan to GMI's team.