Your node. Connected.

The EIZO agent runs alongside your compute tools. It connects outbound to the master, reports system health, relays permitted operations, and returns completed media. The browser controls the server; the worker coordinates the durable queues; agents do the work on each machine.

Enroll a node

1. Install the agent binary and provision its dedicated service account and configuration. RPM packages and the public downloads page are planned; this release uses the existing operator-managed installation.
2. Open Infrastructure and choose Enroll a node. Set a unique name and select only the capabilities this machine is prepared to provide.
3. Save the one-time token and transfer it to the node through your trusted administrative connection. Give the file mode 0600, owned by the account that runs the agent.
4. Configure the same node_name in /etc/eizo/agent.toml. Provision trusted master URLs and local file roots.
5. Run enrollment as the dedicated agent account, then remove the consumed token file.

/opt/eizo-agent/current --config /etc/eizo/agent.toml enrol \
  --token-file /path/to/eizo-enrollment.token

Tokens expire in at most 60 minutes and can be used once. Revoking an unused enrollment in Infrastructure prevents it from connecting. Enrollment creates the node; its first heartbeat provides live inventory.

Run and inspect the agent

With the service unit installed and permissions configured by the operator:

sudo systemctl enable --now eizo-agent
sudo systemctl status eizo-agent
sudo journalctl -u eizo-agent -n 50

The agent reports CPU, memory, GPUs, disks, ComfyUI readiness, and connection-route status. You can inspect a node from Infrastructure or print inventory locally:

/opt/eizo-agent/current inventory
/opt/eizo-agent/current --config /etc/eizo/agent.toml check

An inventory-only run reports health without accepting work:

/opt/eizo-agent/current --config /etc/eizo/agent.toml run --inventory-only

Configure multiple trusted master routes for failover. Service identity and signing trust stay pinned; arbitrary discovered endpoints do not become trusted automatically.

Render, collect, and transfer

The worker assigns work only when a node is eligible. The agent submits the workflow to local ComfyUI and tracks its prompt. Completion comes from ComfyUI history, rather than guessing from a newly appearing file. The output journal survives agent restarts and retains work until the master confirms stored artifacts.

Files use bounded, resumable HTTPS transfers with content hashes. Explicit file operations can also use locally configured SSH/SCP peers. Cross-account render transfers use scoped HTTPS access so the agent can read only the assigned render’s declared input blobs and publish into the submitting account’s storage.

Pause stops new assignments while letting assigned work finish. Offline or stale nodes are not newly allocated work. An uncertain execution is retained for reconciliation instead of being blindly resubmitted.

Personal, shared, and leased compute

New nodes are private to their owner. Explicit account grants permit shared use. Public leasing is opt-in and requires approved workflow templates. A node needs an agent that reports scoped-execution support before cross-account work is eligible.

Funded time reservations receive the next available slot before personal work. They never preempt an already running render. The lease window starts at allocation; after expiry, assigned work may finish but no new leased jobs start. The submitting account’s storage remains separate from the node owner’s storage.

Update safely

The update system verifies signed release metadata, version, expiry, hash, and length before activation. Node identity and configuration survive an upgrade. Updates drain work before switching binaries and support rollback. Agent binaries and permitted local maintenance commands are managed separately from application login credentials.

ComfyUI maintenance uses commands explicitly provisioned on that node. The agent does not accept arbitrary shell instructions from the browser. Do not replace enrolled identity state as part of a routine update.

Check the right layer

• Offline: check the agent service, trusted routes, TLS, host clock, and network access. The console marks heartbeats older than 60 seconds unavailable.
• Connected but not ready: check local ComfyUI activity, draining state, and allowed render capabilities. Inventory-only agents do not accept work.
• Waiting render: check node access, inputs, dependencies, supported workflows, and any lease window.
• Outputs pending: keep the agent’s state directory intact. Its collector retries upload and registration until the master acknowledges storage.
• Enrollment rejected: check the token’s name binding and expiry. Create a fresh token if the original expired or was revoked.

Back to Help →