Bare-metal seller guide

Outcome. A bare-metal storefront that publishes whole-machine SSH-access listings, grants a buyer tenant access after settlement, and reclaims that access when the lease ends.

Before you run

The examples target Base Sepolia (chain ID 84532) with test USDC and require a physical machine the provisioning service can reach over SSH.

The source guide's git clone follows moving repository state. Immediately after cloning and before building, select the inspected revision:

git checkout --detach 648682c67a26caf9283492e999923ab6ca1206ee

---

How to bring up a bare-metal storefront: publish whole-machine SSH access listings, grant a buyer SSH access after settlement, and reclaim that access when the lease ends.

The operator model is one storefront per domain. To sell VM slices and bare-metal access for the same physical host, run a VM storefront and a bare-metal storefront against the same provisioning/site authority. Use the same attribute.physical_host_id in both storefront inventories so shared host accounting can prevent double selling.

For VM slices, see seller-quickstart.md. To run your own listing registry instead of pointing at an existing one, see indexer-quickstart.md.

Transitional packaging note: the bare-metal domain schema lives in domains/bare_metal, while the current runnable provisioning/site authority still lives under provisioning/compute/service. Until the storefront package split is complete, some compose service names and commands still use the VM seller image.

Prerequisites

• Linux host with Docker + docker compose v2.
• A wallet on the EVM chain you'll operate on, funded with gas plus whatever ERC-20 you'll accept as payment. The examples use Base Sepolia + USDC at 0x036CbD53842c5426634e7929541eC2318f3dCF7e.
• An RPC URL for that chain.
• A listing registry URL + write token if the registry gates writes.
• A physical machine that the provisioning service can reach over SSH.
• An SSH keypair for the provisioning container, with the public key installed on each bare-metal node and passwordless sudo enabled for the Ansible user.

1. Get the code and build

git clone https://github.com/arkhai-io/simple-compute-market.git
cd simple-compute-market
make build-seller

build-seller currently builds the shared seller images used by the transitional VM and bare-metal paths.

At this revision

Read the transitional packaging as current behavior, not a roadmap: the bare-metal domain schema is already split out under domains/bare_metal, while this build and some compose services still use the VM seller image.

2. Configure

The storefront reads /etc/arkhai/storefront.toml inside the container, which the compose mounts from ./config.seller.toml:

agent_id         = "bare_metal_seller_one"

port             = 8001
base_url         = "http://<YOUR_PUBLIC_IP>:8001/"

db_path          = "./src/market_storefront/data/storefront/agent.db"
log_file_path    = "./logs/seller.log"
admin_api_key    = "<choose-a-secret>"

[wallet]
private_key    = "0x<YOUR_SELLER_PRIVATE_KEY>"
ssh_public_key = "ssh-ed25519 AAAA...placeholder seller@host"

[chains.base_sepolia]
chain_id = 84532
rpc_url  = "https://sepolia.base.org"

[registry]
urls = ["http://34.41.205.175/registry"]

[registry.auth]
"http://34.41.205.175/registry" = "<your-write-token>"

[provisioning]
service_url = "http://seller-provisioning:8081"
mode        = "http"

[pricing]
default_min_price            = "10"
default_token_address        = "0x036CbD53842c5426634e7929541eC2318f3dCF7e"
default_max_duration_seconds = 86400

For co-selling with a VM storefront, use a distinct agent_id, base_url, port, database, and wallet identity as appropriate for the bare-metal storefront. Both storefronts can point at the same provisioning service.

3. resources.csv

Use one resources.csv row per whole machine:

resource_id,resource_type,resource_subtype,unit,value,state,min_price,token,max_duration_seconds,attribute.gpu_model,attribute.sla,attribute.region,attribute.vm_host,attribute.physical_host_id,attribute.allocation_mode,attribute.machine_id,attribute.vcpu_count,attribute.ram_gb,attribute.disk_gb,attribute.virtualization_type
whole-host-001,compute.gpu,H200,count,8,available,10,0x036CbD53842c5426634e7929541eC2318f3dCF7e,86400,H200,99.0,"California, US",,host-ca-h200-01,exclusive,bm-host-ca-h200-01,192,2048,20000,bare_metal

Important fields:

• attribute.allocation_mode = exclusive marks the row as a whole-host listing candidate.
• attribute.physical_host_id is the stable physical identity used for cross-domain accounting. VM slice rows for the same host must use the same value in the VM storefront inventory.
• attribute.machine_id is the provisioning executor's bare-metal node alias. It must match [bare_metal_nodes] in the provisioning inventory. If omitted, resource_id is used.
• attribute.vm_host is for VM slices; leave it empty in bare-metal rows.
• value is the total units on the physical host. For GPU hosts, this is normally the total GPU count.

4. Add the bare-metal inventory group

Edit the provisioning inventory and add each bare-metal node under [bare_metal_nodes]:

cd domains/vms/provisioning/iac/ansible/inventory
cp hosts.example hosts
# edit hosts with your real bare-metal node(s)

[bare_metal_nodes]
bm-host-ca-h200-01  ansible_host=10.0.0.25  public_host=203.0.113.25  ansible_user=ubuntu  ansible_ssh_private_key_file=~/.ssh/id_ed25519

The alias (bm-host-ca-h200-01) must match attribute.machine_id. ansible_host is how the provisioning service reaches the node. public_host is optional; set it when buyers should SSH to a different address than the one Ansible uses. Rebuild after editing the baked inventory:

make build-seller

5. Choose reclaim behavior

Bare-metal grant installs the buyer's SSH public key for a tenant account. On release, the provisioning service runs node_reclaim_access using one of these policies:

• remove_lease_key removes only the SSH key recorded for the lease.
• lock_user removes the lease key and locks the tenant account.
• delete_user deletes the tenant account and home directory.

The default is remove_lease_key. Override with provisioning config:

bare_metal_reclaim_policy: "lock_user"

or with an environment variable:

PROVISIONING_BARE_METAL_RECLAIM_POLICY=lock_user

Use delete_user only for machines where tenant home directories are expected to be disposable.

6. Bring it up

Warning

This brings up the live stack: after settlement, the provisioning service grants — and at lease end reclaims — real SSH access on the physical machines in your [bare_metal_nodes] inventory. Point it only at machines you operate, and note that the delete_user reclaim policy from §5 deletes the tenant account and home directory when a lease ends.

SELLER_CONFIG_PATH="$PWD/config.seller.toml" \
SELLER_RESOURCES_CSV="$PWD/resources.csv" \
SELLER_SSH_PRIVKEY="$PWD/keys/id_ed25519" \
docker compose -f compose/seller.yml -f compose/seller.live.yml up -d

For co-selling, run the VM and bare-metal storefronts as separate deployments with separate configs and public ports, while pointing both at the same provisioning/site authority.

7. Validate the live node

Before publishing, check that the provisioning container can reach the bare-metal node:

docker compose -f compose/seller.yml -f compose/seller.live.yml exec \
  seller-provisioning ansible \
  -i /opt/domains/vms/provisioning/iac/ansible/inventory/hosts \
  bm-host-ca-h200-01 -m ping

Then verify the host is registered and enabled in the provisioning service:

docker compose -f compose/seller.yml -f compose/seller.live.yml exec \
  seller-provisioning curl -s \
  -H "X-Admin-Key: <admin_api_key>" \
  http://localhost:8081/api/v1/hosts/bm-host-ca-h200-01 | jq .

Bare-metal grant/reclaim refuses to queue work for a missing or disabled machine.

8. Publish and inspect

Warning

The publish request is authenticated by the seller wallet and the listing is attributed to that signing address; this is not real-world identity verification. This guide's configuration preserves the source's named public preproduction Registry URL and write-token example, which this website verified only as an optional public read target — publication authority was neither tested nor granted. Before publishing, substitute a Registry you operate or a write endpoint and credential its operator explicitly gave you.

docker compose -f compose/seller.yml exec seller-storefront \
  market-storefront publish --inventory /app/resources.csv

Bare-metal listing payloads have kind = "bare_metal.v1" and include machine_id, physical_host_id, access_method = "ssh", and the advertised duration/price constraints. Inspect local listings:

curl -s http://<YOUR_PUBLIC_IP>:8001/api/v1/listings \
  | jq '.listings[] | select(.offer_resource.kind == "bare_metal.v1")'

Operational notes

• Do not use the same machine as a general-purpose operator login and a bare-metal tenant target unless the tenant account is tightly isolated.
• remove_lease_key is safest for preserving tenant data; lock_user and delete_user are stronger cleanup actions with more operational blast radius.
• Keep physical_host_id stable across CSV edits. Changing it breaks cross-domain accounting for existing rows.
• Keep machine_id stable while leases are active. It is the executor target used for grant/reclaim.
• Re-run publish after importing resource or capacity changes so stale listings close and newly available listings reopen.

Common pitfalls

• attribute.machine_id must match [bare_metal_nodes]. Wrong alias = grant/reclaim fails before Ansible runs.
• Bare-metal and VM rows for one physical host must share attribute.physical_host_id. Otherwise the site ledger cannot prevent cross-mode double selling.
• Do not put VM slice and bare-metal rows in the same storefront inventory once the storefront split is complete. Run one storefront per domain and let the shared site authority coordinate host accounting.
• agent_id must be a Python identifier — no dashes.