AI & IT toolsAI Workflow

Using Higgsfield from Claude Code: connect over MCP, get a quote first, submit one job at a time, and keep a log

Compare the MCP and CLI ways of connecting Higgsfield to Claude Code, then follow the real order we used for one scene of a sponsored video: checking the connection, uploading files, making a style frame, getting a quote, submitting the video, waiting for the result, and keeping a credit ledger.

AZTechnology13 min read

한국어English

Based on: the official Claude Code MCP docs, the Higgsfield MCP and CLI help pages (checked on November 4, 2026), and our work log from October 2026. Images are official help pages, one frame from a Claude Code terminal recording, and stills from a video we made. The plan and remaining-credit cells in the terminal image are blurred. The terminal text is in Korean. AZTechnology was once paid by Higgsfield to make a sponsored video. That sponsorship has ended, and this post was written without any request or review from Higgsfield. The result stills and work log here come from making that sponsored video.

Once Higgsfield is connected to Claude Code, you can make images and videos by chatting, without opening the website. Uploading a Blender render, making a style frame, asking the price, submitting the video, and downloading and checking the result all happen in one terminal. With several scenes, repeating the same steps is easy too.

There is one thing to watch. Anything made through an agent always takes credits, whatever your plan. Because you give instructions in plain language, one misunderstanding can spend credits for nothing. So this post covers safeguards for your credits along with how to connect.

After this post you can:

  1. Connect Higgsfield to Claude Code over MCP or the CLI.
  2. Check the connection by asking in chat.
  3. Have Claude Code get a quote first and submit one job at a time.
  4. Log the cost and job number for every generation.

Before you start: Claude Code must be installed (Claude Code installation). You need a Higgsfield account and credits. Read Reading Higgsfield pricing and credits first for prices. If MCP is new to you, start with What is MCP.

1. MCP and CLI: two ways to connect

MCP is an open protocol that lets an AI agent call the features of an outside service. In restaurant terms, it hands the whole menu to the customer (the agent), who can pick any dish and any option. CLI and Skills are set meals for orders you place often. There is less to choose, but the order is short and the results are consistent.

Higgsfield offers both.

Higgsfield help page How do I access Higgsfield via CLI?, with the When should I use CLI instead of MCP? table (rows Best for, Token overhead, Output format, Credits, Unlimited models, API key required) and the three steps of How do I set up CLI + Skills? The official help table comparing CLI and MCP, and the three CLI setup steps (Higgsfield help page screenshot)

CompareCLI + SkillsMCP
Best forCoding agents (Claude Code, Cursor, Codex)Chat agents (Claude on the web and others)
Token useLowThe full tool description on each call
OutputShaped by the skillThe agent picks options freely
CreditsSame account, same ratesSame account, same rates
Unlimited modelsNoNo
API keyNot neededNot needed

By that table, the CLI fits coding agents such as Claude Code. For the sponsored video, we attached MCP to Claude Code instead. We had to set options such as the video edit mode, resolution, sound and quoting ourselves for each scene. We did not use the CLI for this work.

Higgsfield help page What is Higgsfield MCP, and how is it different from the web?, showing the MCP connector URL and a table comparing Web (higgsfield.ai) and MCP on Unlimited model access, Free generations and Credit cost The official help table comparing the web and MCP. In the MCP column, Unlimited model access says "No, always deducts credits" (Higgsfield help page screenshot)

The table above also sums up how MCP differs from the web. Models and prices are the same, and only the web gets unlimited models and free generations. Results made over MCP also land in your Assets list on the website, tagged as MCP.

2. Connecting

Pick one of three ways. The server address is https://mcp.higgsfield.ai/mcp in every case, and you log in through the browser with no API key (OAuth).

A. Add it as a claude.ai connector (the way we used)

If you use Claude Code logged in with a claude.ai account, connectors you add on claude.ai show up in Claude Code automatically.

  1. In a browser, open the connector settings on claude.ai (claude.ai/customize/connectors).
  2. Add Higgsfield and log in with your Higgsfield account when asked.
  3. In Claude Code, type /mcp. Higgsfield should appear in the list as a server from claude.ai.

On Team and Enterprise plans, only admins can add connectors.

B. Add it to Claude Code directly

Run the command below in a terminal. It is Claude Code's standard form for adding a remote MCP server, with the Higgsfield server address filled in. We used way A for our work, so we did not run this command.

claude mcp add --transport http higgsfield https://mcp.higgsfield.ai/mcp

Open Claude Code, type /mcp, choose Higgsfield, and log in through the browser. To use it in every project, add --scope user. The default is local, which applies to the current project only.

C. Use the CLI and Skills

  1. Run npx skills add higgsfield-ai/skills. Three skills are installed (generate, soul, product-photoshoot).
  2. Log in with higgsfield auth login.
  3. In Claude Code, call /higgsfield:generate or say something like "Generate an image with Higgsfield."

3. Checking the connection

Once connected, ask for the status before you make anything. Checking the status and the balance costs no credits.

Do this: Send something like this to Claude Code.

Check the Higgsfield MCP connection. Don't generate anything; just read my account and remaining credits.

A Claude Code terminal in Korean. The answer to a question about the Higgsfield MCP connection, with a table showing the connection is working; then, after asking whether a Blender scene can be made as a blockout and turned into video with Higgsfield, the line Called blender, Higgsfield 2 times and a table of the Blender MCP and Higgsfield MCP connection status Checking the connection in Claude Code while recording the sponsored video. The plan and remaining-credit cells are blurred (one frame from a Claude Code terminal recording)

The image shows Blender MCP and Higgsfield MCP attached together. For the second question, Claude Code read the Blender scene (object count, camera, light) and the Higgsfield account separately and answered with a table. Once you see this table, you can go from blockout to conversion in one conversation. Attaching Blender is covered in Claude Code and the Blender MCP.

4. Handing over one scene, step by step

Below is the actual order in which we ran one car chase scene through Claude Code. The tool calls are copied in order from the session log.

StepWhat Claude Code calledResultCredits
1media_upload (source MP4, still at 1.5 s)Two files uploaded0
2generate_image (gpt_image_2_5, "Edit Image 1…")One style frame1.5
3A person approved "submit" after seeing the estimated price (about 76)
4Pre-submission check script24 fps source, one style frame, per-scene cap, balance floor, approval record all passed0
5generate_video (seedance_2_5, video_edit, 720p, get_cost: true)Quote of 76 credits, nothing generated0
6generate_video (same values, submitted)Job number, pending76
7jobs_wait (job number)Done after about 6 minutes0
8Download → frozen frame check4 frozen frames0
9Fill frozen frames in editing0 frozen frames0

Step 5 is the key. Adding get_cost: true to the same tool returns only the price without generating anything. It is the same as reading the number on the Generate button on the web. The person approved based on the estimated price, and right before submitting, Claude Code checked once more that the real quote matched before moving to step 6.

Top, a still of a night street blockout made of grey and blue boxes. Bottom, a style frame that turns the same framing into a rainy night live-action picture Top, the blockout still uploaded in step 1. Bottom, the style frame made in step 2 (Blender render; bottom is a Higgsfield result)

Two moments of a silver-grey sedan and a red coupe driving through a rainy night city. Left at 1.5 s, right at about 7.5 s, with the red car moved from front left to the right Stills at 1.5 s and about 7.5 s from the result video received in step 6 (Higgsfield result, sponsored video material)

Do this: When you try this with your own scene, do not ask for everything at once. Send it in stages as below. These requests are rewritten from our work; we did not run this exact text. The answers and call order will differ a little each time.

1) Upload source.mp4 and still.png to Higgsfield and just tell me the result. Don't generate anything.
2) Turn still.png into one live-action style frame with GPT Image 2.5. Tell me the price first and make it only after I approve.
3) Get only a quote (get_cost) for converting source.mp4 with Seedance 2.5 video edit, 720p, sound on. Don't submit.
4) (After seeing the price) Submit. When it's done, download the result to this folder and log the job number and actual charge in the ledger.

5. Write the credit safeguards down as rules

A tool you instruct in plain language accepts "try again" far too easily. So we wrote the following into the project folder's rules file (CLAUDE.md) before handing over the work. Claude Code reads this file first in every session.

SafeguardThe value we set
Quote firstGet the price with get_cost before submitting and write it in the log file
One job at a timeSubmit the next job only after reviewing the result
Per-scene capUp to three paid 720p videos per scene
Balance floorStop and report if remaining credits drop below 800
Human approvalGenerate video only after a person says "submit"
Pre-submission checkA script checks for a 24 fps source, one style frame, and the balance and quote

Do this: Add a few lines like these to CLAUDE.md in your work folder.

## Higgsfield rules
- Before making any image or video, give me the price with get_cost and wait for my approval.
- Submit only one job at a time.
- If remaining credits drop below ○○, stop and ask.
- For each generation, log the date, model, settings, quote, actual charge and job number on one line in cost.md.

6. Keeping a ledger for every generation

For every generation, we kept the fields below in the log file.

FieldExample
Date, scene, stage10-04, car chase, 720p
Model and settingsseedance_2_5, video_edit, draft off
SourceFile name and sha256 (file fingerprint)
Quote76
Actual charge76 (checked in the transaction history)
Remaining creditsThe balance lookup value
Job numberUsed to find it in the web library

sha256 is a long fingerprint calculated from a file's contents. When you fix a file and resubmit it under the same name, it tells you which version you sent. We checked the actual charges with the Higgsfield MCP transaction history (transactions) and balance lookups. At the end, we looked up the balance and compared it with the ledger's calculated value, and the difference was 0.

7. Web or MCP: when to use which

WhenChoose
Trying a few things on a plan with unlimited modelsThe web (unlimited works only there)
Many scenes where you repeat the same stepsClaude Code + MCP
One flow from Blender render to conversion and checksClaude Code + Blender MCP + Higgsfield MCP
Repeating a fixed format (such as product photos) with few tokensClaude Code + CLI and Skills
Recording the generation screen to show in a videoThe web (MCP returns only the result, so the process is not on screen)

Using Higgsfield over MCP or the CLI falls under the Developer Terms in the terms of use. Everything an agent makes with your account is your responsibility. Use the connection only in folders and sessions you trust, and when you are done, you can clear the login from the /mcp menu.

Practice

  1. Attach Higgsfield in one of the three ways from section 2 and check that it appears in the /mcp list.
  2. Ask for the connection status with the request in section 3. Compare with the balance on the website to confirm no credits were used.
  3. Add the four Higgsfield rule lines from section 5 to CLAUDE.md in your work folder, then ask for the price of one image. Check that Claude Code stops and waits for approval before making it.

Common mistakes and fixes

SymptomCauseFix
Higgsfield is not in the /mcp listThe connector was added to a different account, or you have not logged in yetCheck which claude.ai account Claude Code is logged in with, and log in again
An unlimited model still took creditsUnlimited does not apply to MCP or CLI generationsMake unlimited items on the website
You only asked for a price, but it generated anywayThe request did not say "don't make it"Ask for get_cost only, and put an approval step in the rules file
The same scene was submitted several timesSeveral jobs were handed over at onceOne job at a time, with a per-scene cap
You cannot tell which version you submittedFiles were overwritten without a logLog the sha256 and job number for every generation
The result does not show up on the websiteYou are looking at a different list in the libraryLook for items tagged MCP in Assets on the website

In one line

Connect with a connector, claude mcp add, or the CLI, and before making anything always go quote (get_cost) → human approval → one submission → ledger entry.

Higgsfield posts: Higgsfield introduction · Reading pricing and credits · Blockout to live-action video

References

KEEP EXPLORING

Read next

Bar chart. The status board grew from 65 lines on October 3 to 114 on the 9th, 131 on the 10th and 143 on the 11th, then stayed at 80 to 83 lines from October 12 to November 4AI & IT tools
AI Workflow17 min read

Keeping a multi-week Claude Code project on track: RESUME files, handoffs and context limits

A long Claude Code chat gets summarized or cut as it grows, so for work that runs for weeks we split it into sessions with a short 'current summary' in a RESUME file, a /handoff command, and a status line and hook that report context usage. We follow the real records of our working folder and an example we ran ourselves.

Claude Code Docs page. In the left menu, Automate actions with hooks is selected under Automation. The centre shows the same title and a description starting with Hooks are user-defined shell commandsAI & IT tools
AI Workflow16 min read

Claude Code hooks: blocking risky agent commands with a PreToolUse guard

Build a small Python script that Claude Code runs just before an AI agent uses a tool, and use it to block five things: deleting lock files, wrapping Blender in timeout, factory-startup installs, running Blender without the lock script, and reading secret files. Real test output included.

Four frames from the first version of the air defense game. Launchers in a low-poly desert at sunset, the score, base health 100, and a green bar at the top left, time left 0:30, 0:17, 0:07, and 0:00 at the top right, and in the last frame the Korean text for Defense succeeded! Score 1500, R: restart (Godot run screenshot)AI & IT tools
AI Workflow13 min read

Setting the scope of your first game: a size a solo developer can finish, and how to schedule it

To take a first solo game all the way to the end, you need to cut it down to the right size. This guide covers what to write on a one-page plan and how to split the schedule into steps, with real times, snags, and scope changes from making three small games in Godot.

Air defense game screen in a low-poly desert at sunset. Box 1 is the on-screen text at the top left, box 2 is missile smoke and an explosion, box 3 is a launcher truck at the bottom left (Godot run screenshot)AI & IT tools
AI Workflow18 min read

Splitting solo game dev work with AI: what to hand off, what to check, terms, and Steam and itch.io AI disclosure

When you make a game alone, how much of the code, 3D assets, art, sound, and text can you hand to AI, and where must a person check the result? This guide covers the commercial-use terms of each kind of tool, the AI disclosure rules on Steam and itch.io, and what AI actually did in a Godot game test.

A low-poly desert at sunset with four launchers. A missile rises from the left launcher, a three-ring aim circle with a yellow 90% sits among red diamond markers in the sky, and the top left shows remaining launchers 4/4 and remaining missiles 31 (Godot run screenshot)AI & IT tools
AI Workflow13 min read

A map of solo game development: what Blender, AI, and a game engine each do, and the steps to your first game

When you make a game alone, Blender, an AI assistant, and a game engine each take a different part of the work. This guide sorts out who does what, lays out seven steps to a first finished game, and shows notes from making small games in Godot. It is the starting page of a five-part series on solo development.

The Script screen of the Godot editor. Under func _steer at line 102 of missile.gd are var to_t := target.position - position, var dist := to_t.length(), var want := (lead - position).normalized(), and var ang := dir.angle_to(want), with a list of functions on the leftAI & IT tools
AI Workflow16 min read

What helps to know before solo game development: what you need without a degree and what AI fills in

When you make a game alone, which parts of programming, math (vectors and coordinates), art, sound, and game design do you really need, and which can AI and tools fill in? This guide answers with real game code and notes, including what AI cannot fill in, and points to the official learning material for each engine.