@assemblyline-agents/higgsfield

Official Assembly Line connection plugin for Higgsfield's CLI-first coding-agent workflow. Higgsfield remains a normal filesystem-declared connection; its transport runs the official CLI inside the active agent sandbox rather than launching a process on the Assembly Line gateway.

Install and configure

assembly-line add higgsfield agent

This creates connections/higgsfield.ts:

import { defineHiggsfieldConnection } from "@assemblyline-agents/higgsfield";

export default defineHiggsfieldConnection({
  // Reviewed tools are enabled by default.
});

Set access: "approval-required" when generation and editing should pause for approval.

Use a read-only policy when the agent should inspect Higgsfield without generating or editing:

export default defineHiggsfieldConnection({ access: "read-only" });

The connection exposes higgsfield__read for discovery, status, cost, history, and polling, plus higgsfield__write for upload, generation, training, editing, reframing, and assembly. The write operation can create remote assets and consume account credits.

Assign it to a subagent with the same connection abstraction:

import { defineAgent, useModel } from "@assemblyline-agents/core";

export default defineAgent({
  description: "Create and edit social video.",
  connections: ["higgsfield"],
  setup() {
    useModel("openai/gpt-5.4-mini");
  }
});

Why the transport is a sandbox CLI

The connection definition compiles as protocol: "cli" and transport: "sandbox". Tool discovery, subagent scoping, access policy, approval, tracing, and filesystem activation stay in the connection layer. At execution time, Assembly Line invokes the reviewed command in the active run sandbox, where it can consume /files/... attachments and write results to /workspace.

Model-authored shell source is never accepted. The plugin validates the official command surface, Assembly Line shell-quotes each argument, hydrates logical input paths, and adds --json --no-color.

Sandbox setup

Install the official CLI in the sandbox image or template:

curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh

Then authenticate interactively inside the persistent sandbox belonging to that user:

higgsfield auth login

Do not bake CLI credentials into a shared image or reuse one authenticated sandbox across unrelated users. Higgsfield's current CLI login is browser-based OAuth with a loopback callback, so sandbox provisioning must make that login reachable. If the sandbox is ephemeral, retain the user's authenticated CLI home with a user-scoped snapshot or persistent volume.

The plugin rejects auth-token, website repository-token, and secret-management commands.

Official references:

• https://higgsfield.ai/cli
• https://github.com/higgsfield-ai/cli