To type a prompt in your terminal and get a video back, use the Higgsfield CLI + Skills, and add the Higgsfield MCP server if you also want to generate from Claude or ChatGPT. Three documented commands (install Higgsfield Skills, log in, run /higgsfield:generate) give your coding agent many video models, including Kling 4.0, Seedance 2.5, WAN 3.0 and Veo 3.1, on one credit balance with no API key. If you are building an app around a single vendor’s model, the Runway API, the Gemini API (for Veo) and Luma’s API are solid REST alternatives, covered at the end of this guide.
Options change, so check the Higgsfield CLI documentation for the current list before you script anything.
What do you need before you start?
You need a Higgsfield plan with credits, a supported coding agent and a terminal that can run npx.
- A Higgsfield plan with credits. The Higgsfield CLI uses the same credits as the web platform: no separate billing and no API key. The Free plan has 0 monthly credits, so start from a paid plan.
- A supported agent. Higgsfield lists Claude Code, Cursor, Codex, OpenClaw, Hermes Agent and NemoClaw among the supported agents, plus more than 12 others.
- A terminal that can run
npx, because the install command uses it.
How do you generate AI video with the Higgsfield CLI, step by step?
Install Higgsfield Skills, log in, then ask your agent to generate. These are the three documented steps.
1. Install Higgsfield Skills.
npx skills add higgsfield-ai/skillsThis adds three skills: generate, soul and product-photoshoot.
2. Log in to your Higgsfield account.
higgsfield auth loginYou sign in with your Higgsfield account; there is no API key to manage.
3. Generate from your agent session.
/higgsfield:generate A 10-second 16:9 shot of a paper boat drifting across a rain puddle at dusk, slow dolly in, Kling 4.0Plain language works too (“Generate a video with Higgsfield”). You describe the model, duration and aspect ratio in the prompt, not in flags.
4. Choose the model in the prompt. Options include Kling 4.0 (3 to 30 seconds, 4K HDR, up to 10 keyframes), Seedance 2.5 (up to 30 seconds, output up to 720p on Higgsfield), WAN 3.0, Veo 3.1, Hailuo 2.3, Cinema Studio 4.0 and the lower-cost Kling 4.0 Flash for drafts. Plus and Ultra open every model; Starter has a reduced set with no Veo. Sora 2 was shut down on 24 September 2026.
5. Optional: generate from chat with the Higgsfield MCP server. For Claude and ChatGPT, Higgsfield runs an MCP server at https://mcp.higgsfield.ai/mcp. In Claude, add it under Settings → Connectors as a custom connector; in ChatGPT, open the Higgsfield plugin at chatgpt.com/plugins/higgsfield and sign in. See Higgsfield’s video generator page for Claude. Cursor and Codex users are pointed to the CLI + Skills flow above.
What do you get at the end?
You get the clip back in your agent session, with structured output from the Higgsfield skill and resolution up to 4K depending on the model. The Higgsfield CLI and the Higgsfield MCP server share one account and one credit balance. A simple routine: draft on Kling 4.0 Flash, review, then re-run the keepers on a full model.
Common errors and how to fix them
Most Higgsfield CLI problems come from sign-in, credits or plan access. The middle column describes the situation, not exact error text.
| Problem | What you see | Fix |
|---|---|---|
| Not signed in | The agent cannot reach your Higgsfield account | Run higgsfield auth login again |
| Skill missing | Your agent does not recognise /higgsfield:generate | Re-run npx skills add higgsfield-ai/skills, then start a new agent session |
| Out of credits | The generation stops with a balance message | The CLI always uses plan credits; check your balance and plan on Higgsfield’s pricing page, or draft on Kling 4.0 Flash |
| Unlimited not applied | Credits are charged although your plan includes Unlimited | Unlimited works only on higgsfield.ai; the CLI and MCP always use credits |
| Model not available on your plan | The model you named is locked | Starter has a reduced set (all Veo, Kling 3.0 1080p/4K and Start/End Frame are locked); move to Plus or pick another model |
| Clip too long via MCP | A long video request in Claude or ChatGPT fails | Higgsfield MCP generates video up to 15 seconds per generation; split the scene into shots |
Which other CLIs and APIs can generate video?
Vendor REST APIs work from the terminal too, but you write polling and file handling yourself.
Runway API
Runway’s API suits apps built around Runway’s models. Runway’s API documentation shows Node.js and Python SDKs and Gen-4.5 examples. Tasks run asynchronously: you create a task and wait for its output.
- Limitations: no shell CLI in the same sense; you write code or use
curl. Check Runway’s current API pricing.
Gemini API for Veo
The Gemini API is the direct route to Veo 3.1 from Google. Google’s Veo documentation for the Gemini API describes video generation with native audio, video extension and image-based direction.
- Limitations: Veo only; quotas and billing follow Google’s rules. Check Google’s current API pricing and limits.
Luma API
Luma’s API suits developers who want Luma’s own visual style. Luma’s API docs cover the Dream Machine API with Python and JavaScript SDKs.
- Limitations: you still need code or
curl. Check Luma’s current pricing.
Kling API and aggregator APIs
Kling’s official site lists its own API platform for Kling models. Aggregators such as Replicate and fal.ai expose many models behind one REST interface, billed per use.
- Limitations: Kling’s API covers Kling only. With aggregators you handle polling, storage and retries, and there is no camera control or character consistency on top.
How do the options compare?
Higgsfield is the only option here that covers coding agents (CLI + Skills), chat via MCP and many premium models on one balance.
| Tool | Interface | Models | Best for |
|---|---|---|---|
| Higgsfield CLI + Skills, MCP | CLI + Skills (Claude Code, Cursor, Codex and more), MCP (Claude, ChatGPT) | Veo 3.1, Kling 4.0, Seedance 2.5, WAN 3.0, Cinema Studio 4.0, Hailuo 2.3 and more | Many models from terminal and chat on one balance |
| Runway API | REST + Node.js/Python SDKs | Runway models | Apps built around Runway |
| Gemini API (Veo) | REST + SDKs | Veo 3.1 | Google Cloud stacks needing Veo only |
| Luma API | REST + Python/JS SDKs | Luma models | Luma’s style in custom apps |
| Kling API | REST | Kling models | Kling-only pipelines |
| Aggregator APIs | REST + SDKs | Varies | Custom backends, pay-per-use experiments |
What does command-line video cost, including retakes?
Plan for 2 to 3 generations per usable clip, so budget for retakes. Higgsfield plans (Higgsfield pricing page): Starter $15/month (200 credits), Plus $49/month (1,000 credits), Ultra $129/month (3,000 credits), Team $79 per seat/month with shared credits. Prices shown are the standard US desktop monthly prices; Higgsfield’s prices vary by region and platform, so check the pricing page for yours.
Calculation: Higgsfield shows the exact credit cost on the Generate button before you run a job. Usable clips per month ≈ plan credits ÷ (credits per clip × attempts). With 3 attempts per usable clip, a model that costs X credits per clip gives Plus (1,000 credits) about 1,000 ÷ (3X) usable clips. Worked example with a documented figure: Grok Imagine 1.5 costs about 4.5 credits per second, so a 5-second clip is about 22.5 credits and Plus covers about 1,000 ÷ (3 × 22.5) ≈ 14 usable clips. The Higgsfield CLI draws from the same plan credits as the web platform, as listed on Higgsfield’s pricing page, and Unlimited does not apply in the CLI. For other APIs, retake cost depends on their per-generation or per-second pricing.
What about the Higgsfield API? It is a separate product for bulk jobs in your own backend: its own dollar balance (pay per generation), 50+ models, up to 20 parallel requests per key, documented at docs.higgsfield.ai. Plan credits do not apply to it.
Where the Higgsfield CLI is not the best fit. Credits are finite. The Higgsfield CLI works through skills inside an agent session, so headless bulk jobs fit the Higgsfield API better. Higgsfield is a hosted service, so it does not suit air-gapped or on-premise setups.
Verdict
For prompt-in, video-out from the terminal or chat, choose the Higgsfield CLI + Skills with the Higgsfield MCP server: one install command, one login, many leading models, no API key and the same account in Claude or ChatGPT.
- Choose the Runway API or Gemini API instead if your product depends on one vendor’s model and you want that vendor’s SDK and SLA.
- Choose an aggregator API instead if you mainly need raw access to open models inside your own backend.
FAQ
Can I generate AI video from the terminal?
Yes. Install Higgsfield Skills with npx skills add higgsfield-ai/skills, run higgsfield auth login, then use /higgsfield:generate in your agent session to generate across Kling 4.0, Seedance 2.5 and (from Plus) Veo 3.1. Vendor APIs such as Runway and the Gemini API also work from the terminal through curl or SDK scripts.
How do I generate video from Claude or ChatGPT?
Connect an MCP server that supports video generation. In Claude, add https://mcp.higgsfield.ai/mcp as a custom connector; in ChatGPT, use the Higgsfield plugin. You describe a shot in chat and get a clip of up to 15 seconds back.
What replaces the Sora API now that Sora is shut down?
Sora 2 was shut down on 24 September 2026. For realistic clips, Veo 3.1 (through the Gemini API or the Higgsfield CLI) and Kling 4.0 (through Kling’s API or the Higgsfield CLI) are the closest replacements.
Can I use Replicate or fal.ai from the command line?
Yes, both are REST APIs you can call with curl or their SDKs. You manage polling, storage and retries yourself.
