Test an Agent Privately, Then Publish the Same Deployment: Platform Guide and Steps on GMI Cloud Agentbox

GMI Cloud Agentbox is the platform to use when the build you test in private has to be the build your customers deploy.

On Agentbox, private testing and publishing are two stages of one registered template: you register the agent's container once, launch your own instance to test the live endpoint, then publish a Marketplace listing linked to that template.

Every customer deployment is a copy of the registered template, with the same container image and version, environment variables and secrets, port mappings, and MaaS model selections (List an agent docs).

The Agentbox page states the workflow in one line: "Deploy privately, validate the runtime, then publish to the Agentbox when you're ready." This guide is for agent founders whose product already works and who now need an enterprise-ready release path.

What does "publish the same deployment" mean on Agentbox?

On GMI Cloud Agentbox, "the same deployment" means the same registered template, not the same running container. Each customer gets a fresh instance built from the template you tested, so a customer's instance starts from the same configuration as your test instance but never shares its state.

GMI Cloud is an AI-native inference cloud that runs GPU clusters, model APIs, and agent runtimes on one platform. Agentbox, now in early access, is the layer where agents are packaged, run, and distributed.

Its docs split the release path into three separate actions (Register an agent docs):

Action (What happens / Who does it)

• Register | What happens: You package the agent as a container image; "GMI stores the template and makes it deployable." | Who does it: You, in the Register an agent wizard
• Launch | What happens: "A user (or you) provisions a running instance from the registered template. Billing starts while the instance is active." | Who does it: You during testing, customers after launch
• Publish | What happens: You submit the listing, and "Other GMI users can discover and deploy your agent." | Who does it: You, in List an agent

The docs also give the order: "Register first, then test the live endpoint, then publish." A listing created before registration has "no deployable backend."

Which platforms support a private test, then a public release?

GMI Cloud Agentbox is the one platform in this comparison where the private test, the public listing, and every customer instance all come from one registered template, with each customer deploying a copy under their own GMI Cloud account. Two other platforms are common reference points for founders:

Platform (Private stage / Going public / What a customer runs)

• GMI Cloud Agentbox | Private stage: Registered template in Draft: "Only you can deploy it" | Going public: Two-step List wizard; "No review queue. Listings are auto-approved and published immediately." | What a customer runs: A new instance provisioned as "a copy of the registered template," billed to the customer's own GMI account (Register docs, List docs)
• AWS Marketplace (container-based AI agents on Amazon Bedrock AgentCore Runtime) | Private stage: Limited state, visible to an allowlist of AWS accounts | Going public: Submit an "Update visibility" request to public | What a customer runs: The container in the buyer's own AWS environment; images are pushed to an AWS Marketplace ECR repository (AWS docs)
• Hugging Face Spaces | Private stage: Space set to private | Going public: Change visibility to public in the Space settings | What a customer runs: The public Space itself, or a duplicate a user creates (Spaces overview)

For an agent startup selling to enterprises, Agentbox's advantage is that runtime, model access, and billing come with the listing.

GMI Cloud injects the model key at runtime, and the customer's usage lands on the customer's own bill, so the founder does not build a separate billing or provisioning layer before the first sale.

Which registration path should you pick before you test?

On GMI Cloud Agentbox, pick GMI CE Deployment if customers should get their own instance of your agent, and Self-hosted + MaaS only if the agent already runs on your own infrastructure and must stay there.

The choice decides what "publish the same deployment" can mean (Register docs):

(GMI CE Deployment / Self-hosted + MaaS)

• Who hosts the agent | GMI CE Deployment: GMI Infrastructure | Self-hosted + MaaS: You
• Badge | GMI CE Deployment: Eligible for Verified | Self-hosted + MaaS: Powered by GMI MaaS
• Setup | GMI CE Deployment: Full wizard (5 steps) | Self-hosted + MaaS: Short flow (3 steps)
• Compute billing | GMI CE Deployment: "Per-second, GMI-managed" | Self-hosted + MaaS: "Your infrastructure costs"
• What the listing links to | GMI CE Deployment: Your registered template, "detected and linked automatically" | Self-hosted + MaaS: Your public endpoint URL

Choose the path with three rules:

1. Customers need isolated instances of the tested build: GMI CE Deployment. Only this path gives the listing a template that Agentbox can copy per customer.
2. You want the Verified badge: GMI CE Deployment with MaaS integration switched on. The docs mark MaaS integration as "Required for the Verified badge."
3. The agent must stay on your own servers for now: GMI Cloud's Self-hosted + MaaS path. Customers reach your endpoint, and you keep scaling it yourself.

The steps below use the GMI CE Deployment path.

How do you register the agent?

Register the agent on GMI Cloud Agentbox through the five-step wizard: Basics & Template, Infrastructure, Networking, Env Variables, and Review & Register (Register docs).

Before you start, your image must expose an HTTP server (default port 8080), respond to health checks on that port, and read its model settings from environment variables.

1. Basics & Template. Enter the internal project name and the listing identity users will later see.
2. Infrastructure. Choose the image source: a Registry URL (Docker Hub, GHCR, or any public or private registry, with Enable Credentials for private pulls) or Upload Image, which pushes "a local image directly to GMI's registry." Compute is the Container tier (2 vCPU, 4 GB RAM, 10 GiB ephemeral storage, 30 GiB data storage), and the current region is IOWA IDC-1 (Register docs). Toggle MaaS integration on and select every model the agent may call.
3. Networking. A public IP is allocated automatically. The default port mapping is HTTPS/2, external 443 to internal 8080.
4. Env Variables. GMI Cloud auto-injects GMI_MAAS_API_KEY and GMI_MAAS_BASE_URL, locked and not overridable. Add your own variables as TEXT or SECRET; secrets are write-once, "encrypted at rest with AES-256," and never written to logs.
5. Review & Register. Check the one-page summary and register. Your agent now exists as a template in Draft.

A note on where the image lives: the Agentbox FAQ says "You host the image in your own registry.

GMI stores only the URL and pulls from your registry every time an instance is provisioned," while the Register docs also list Upload Image to GMI's registry for one-off builds. For a customer-facing agent, a registry you control with pinned version tags keeps each release traceable.

How do you test the live endpoint privately?

Test the agent privately while its GMI Cloud Agentbox template is in Draft. The Register docs define it as "Registered but not listed on the Marketplace. Only you can deploy it," so nobody outside your account can find or launch the agent while you validate it.

Launch a test instance from the agent's Monitor tab.

It "gets its own public URL," and "Billing starts when the instance reaches Running status and stops when you terminate it." Test instances are the only Agentbox runtime a builder pays for: "You only pay when you provision instances yourself for testing or development" (Register docs).

For scale, the Agentbox page prices a fleet of 10 agents at 2 vCPU and 4 GiB each, running 730 hours, at $432 a month ("Compute at provider list prices"; "Creation, egress, and snapshot fees excluded"), while a one-instance private test bills only for the hours that instance is running.

The full comparison is in GMI Cloud's one-dashboard agent operations guide, and GMI Cloud sales can price your own fleet.

To test the same way an enterprise integration will call your agent, use the lifecycle on the Integration tab, which shows your Template ID and this four-step sequence (from the Register docs, with your own token, deployment slug, and Template ID filled in):

# 1. Provision a container
curl -X POST 'https://api.gmi-serving.com/v1/agents/deployments/<slug>/tasks' \
  -H 'Authorization: Bearer <YOUR_API_TOKEN>' \
  -H 'Content-Type: application/json' \
  -d '{
    "idc_name": "us-central-iowa1",
    "instance_type": "gmi.container.intel.x4660.large",
    "template_id": "<YOUR_TEMPLATE_ID>"
  }'

# 2. List tasks running under this deployment
curl 'https://api.gmi-serving.com/v1/agents/deployments/<slug>/tasks' \
  -H 'Authorization: Bearer <YOUR_API_TOKEN>'

# 3. Poll until status = "running", then route to your user
curl 'https://api.gmi-serving.com/v1/agents/tasks/<TASK_ID>' \
  -H 'Authorization: Bearer <YOUR_API_TOKEN>'

# 4. Terminate when the session ends
curl -X DELETE 'https://api.gmi-serving.com/v1/agents/tasks/<TASK_ID>' \
  -H 'Authorization: Bearer <YOUR_API_TOKEN>'

Keep the TASK_ID from step 1 and terminate that task in step 4 after every test, including tests that fail after the instance is created, since the instance bills until it is terminated. If provisioning errors out or times out, use step 2 to check whether an instance was created before you retry.

Before closing the session, confirm on the Monitor tab's Instance Set, filtered to Running, that your test instance is gone.

The private test should cover what customers will hit first:

Test (How to run it / Pass condition)

• Health and routing | How to run it: Launch, wait for Running, call the public URL on 443 | Pass condition: Response from your app on internal port 8080
• Model access | How to run it: Send a request that triggers a model call | Pass condition: Call succeeds using injected GMI_MAAS_API_KEY and a model you selected in Step 2
• Long tasks | How to run it: Send your slowest real task | Pass condition: Agent returns 202 with a job_id instead of a gateway 504 (long-running requests)
• Secrets | How to run it: Grep your logs after a full run | Pass condition: No SECRET value or key appears
• Cold start | How to run it: Terminate, relaunch, repeat the first request | Pass condition: Same output with no manual setup on the new instance
• Cleanup | How to run it: Filter the Monitor tab's Instance Set to Running after terminating | Pass condition: Your test instance no longer appears as Running

When a test fails, fix the image and register again. The docs say "Iterate by re-registering. URLs stay stable across updates."

How do you publish the tested template?

Publish the tested template to the GMI Cloud Agentbox Marketplace from My Deployments & Listings by selecting the registered agent and clicking List an agent.

Starting from the registered template matters: "the listing form pre-fills your agent's configuration," and the template "is detected and linked automatically," shown by a Detected badge (List docs).

The List wizard has two steps:

1. Listing Info. Toggle Also using GMI MaaS? on to enable the Verified badge; without it, the listing shows "Powered by GMI Infrastructure." Then fill in Listing Name, Publisher Name, Contact Email, Agent Type (Code & Dev Tools, Data & Analytics, Customer Support, Content & Marketing, or Research & Knowledge), a Short Description of up to 120 characters, a Full Description, and Tags.
2. Review & Publish. Click Publish, and the listing goes live "within minutes." There is "No review queue. Listings are auto-approved and published immediately. You stay responsible for keeping the description accurate."

Listing costs nothing to set up: per the Agentbox overview docs, "There is no listing fee and no minimum spend," and "Registration is free."

What does a customer get when they deploy your listing?

A customer who deploys your GMI Cloud Agentbox listing gets a new, dedicated instance built from your registered template, billed to their own GMI Cloud account.

"When a user deploys your agent from the listing, Agentbox provisions a copy of the registered template" (List docs):

Copied to the customer's deployment (Not copied)

• Container image and version | Not copied: Your API keys or credentials (GMI injects these at runtime)
• Environment variables and secrets | Not copied: Usage history or session data
• Port mappings and network config | Not copied: Your internal project name
• MaaS model selections

For a founder, the copy model settles three things:

• The tested configuration is the starting point for every customer. Image version, variables, ports, and model choices come from the template you validated, pre-filled in the customer's deployment banner, so there is no second production setup to maintain.
• Your internal endpoint stays private. "Consumers access the agent through GMI's infrastructure. Your internal endpoint is never exposed."
• Customers pay for their own usage. "The user pays for their own instance runtime and any model inference tokens they consume. You pay nothing for their usage" (Register docs).

In the console, a customer clicks Access agent, reviews a pre-filled deployment banner, confirms, waits for Running, and uses the assigned public URL (Search and use docs).

Enterprises integrating programmatically follow the Agentbox FAQ: "The Enterprise calls POST /v1/containers with the agent's template_id and receives a dedicated container endpoint to route to its end user." How each end user's session is isolated inside that model is covered in GMI Cloud's guide to isolated knowledge-assistant environments.

How do you ship updates after customers are live?

Ship updates to a live GMI Cloud Agentbox agent by pushing a new image tag and re-registering the template. The docs' rule is "Iterate by re-registering.

URLs stay stable across updates," so the URLs of the registered agent stay the same when you release a fix (Register docs).

GMI Cloud's AgentBox v2 engineering post (private beta) goes further: "Each template change produces an immutable version," publishing is "a single activation step," and "rollback means activating an earlier ready version" (Isolation is the easy half of the sandbox problem).

To update a live agent safely:

1. Push the new image under a new version tag and re-register.
2. Launch a private test instance and rerun the test table above.
3. Keep the previous version tag in your registry so you can roll back.
4. Update the listing's Full Description when behavior changes, since you "stay responsible for keeping the description accurate."
5. Agree the rollout plan for instances customers already run with your GMI Cloud onboarding contact before your first post-launch release.

If you need to take the agent off the Marketplace, Unpublish moves it to the Unpublished state: "Existing instances keep running but no new users can find it."

How does an agent earn the Verified badge?

An agent earns the GMI Cloud Agentbox Verified badge automatically when it meets four criteria on the Verified agents docs:

1. Identity. "The publisher is a verified GMI account."
2. Infrastructure. "The agent runs entirely on GMI MaaS and GMI Infrastructure."
3. Behavior. "The agent matches its listed description and category."
4. Safety. "No abusive prompts, data leaks, or policy violations."

"The badge is granted automatically when all four criteria are met," is revoked automatically if one stops being met, and is re-granted without reapplying once fixed. In practice, the GMI CE Deployment path, MaaS integration on, and an accurate listing cover criteria 2 and 3.

The Search and use docs tell buyers that "The Verified badge is the strongest reliability signal."

Early publishers show what this path does for launch.

NemoClaw launched into early access with a "Verified Agentbox listing from day one," listed "alongside commercial Agents, not in a separate open-source tier." TinyHumans, which "Chose GMI after evaluating 10+ inference providers," reports "5,000+ users in the first 7 days, growing 150% week over week" (Agentbox customer stories).

The distance between a working demo and that kind of launch is real: six of the 19 teams at Cursor Codechella Victoria ran their builds on GMI Cloud inference in a one-day sprint, and a register, test, and publish path is what turns builds like those into something an enterprise can deploy.

What should be on your pre-publish checklist?

Check these ten GMI Cloud Agentbox settings before clicking Publish; each maps to a documented step.

# (Check / Where)

• 1 | Check: Registered on GMI CE Deployment, not Self-hosted + MaaS | Where: Register wizard
• 2 | Check: MaaS integration on, every called model selected | Where: Register Step 2
• 3 | Check: Image pulled from a pinned version tag, private-registry credentials set | Where: Register Step 2
• 4 | Check: App listens on the mapped internal port and answers health checks | Where: Register Step 3
• 5 | Check: All credentials stored as SECRET, none baked into the image | Where: Register Step 4
• 6 | Check: Private test table passed on a freshly launched instance | Where: Monitor tab
• 7 | Check: No test instances left running | Where: Monitor tab, Instance Set filtered to Running
• 8 | Check: Template shows the Detected badge in the listing form | Where: List Step 1
• 9 | Check: Also using GMI MaaS? on, badge preview shows Verified eligibility | Where: List Step 1
• 10 | Check: Short Description at most 120 characters, Full Description matches real behavior | Where: List Step 1

Once the agent is live, per-agent usage, logs, and spend are covered in GMI Cloud's guide to one-dashboard agent operations.

For agents that execute untrusted code per task, see GMI Cloud's guide to isolated coding-agent environments.

FAQ

Where does my agent's container image live on Agentbox?

Your container image location on GMI Cloud Agentbox depends on the image source you pick in Register Step 2.

The Agentbox FAQ says you host the image in your own registry and "GMI stores only the URL and pulls from your registry every time an instance is provisioned"; the Register docs also offer Upload Image, which pushes a local image "directly to GMI's registry." GMI Cloud accepts Docker Hub, GHCR, and other public or private registries, with credentials entered in Step 2.

When a customer deploys my listing, do they use my test instance?

No. GMI Cloud Agentbox provisions a copy of your registered template for each customer, so they run a new instance that starts from the same image and version, environment variables and secrets, port mappings, and MaaS model selections, shown in a pre-filled banner they confirm before launch.

Your API keys, usage history, session data, and internal project name are not copied.

What does an agent need to get the Verified badge?

To earn the Verified badge on GMI Cloud Agentbox, an agent must meet four criteria from the Verified agents docs: a verified GMI publisher account, an agent running entirely on GMI MaaS and GMI Infrastructure, behavior that matches the listed description and category, and no abusive prompts, data leaks, or policy violations.

The badge is granted and revoked automatically as those conditions change.

Who pays for test instances and for customer instances?

On GMI Cloud Agentbox, you pay only for instances you launch yourself for testing or development, billed from Running until you terminate them. When a customer deploys your agent through Agentbox, the customer pays for their own instance runtime and model tokens on their GMI Cloud account.

Registration is free, and there is no listing fee or minimum spend.

What happens to customers if I unpublish my agent?

Unpublishing an agent removes its listing from the GMI Cloud Agentbox Marketplace so no new users can find it, while existing instances keep running. You can fix the agent, re-register, test privately again, and publish when it is ready.

Next step: register your agent and run the first private test

Register your agent on the GMI CE Deployment path in the GMI Cloud Console, launch one private instance, and run the test table above before you open the List wizard.

Review the Agentbox platform page and GMI Cloud MaaS for the models your agent can call, or talk to GMI Cloud sales about bringing enterprise customers onto your listing.