bInference skill

Teach your agent to check its AI budget before long work, pick models by live price, and read and make images.

A skill is a folder of instructions and small scripts that your agent reads when a task calls for it. The bInference skill works in Claude Code, Codex, OpenClaw and any app that installs skills.

Install

npx skills add binference/skills

The installer asks which of your apps to add it to. The skill reads your key from the environment:

export BINF_API_KEY=binf_...

Its scripts need Node.js 18 or newer. BINF_API_URL changes the API address, which is https://binference.io/api/v1 by default.

npx binference agent-os installs the skill in every agent folder it makes, ready to use. See Set up an agent.

Claude Code also has its own answers about model prices. To make sure it uses the skill, add this line to your project's CLAUDE.md:

This agent runs on bInference. Use the binference skill for AI budget, model prices and choice, sending images to models, and generating images with image models.

What to ask

You ask	The skill
"Watch BNB every 5 minutes for a day. Can my budget take it?"	Prices the whole task on live prices before it starts
"Which cheap model reads images?"	Lists models from the live list, cheapest first
"What does this chart show?"	Sends the image to a model that reads images
"Make a recap card of today's trades"	Makes the image, reads it back and checks every number
"Give the research worker its own key, $2 a day"	Makes a key with that limit over the MCP server
"Why did that call fail with 402?"	Reads the budget and says what to change

Budget check

Before work that makes many model calls, such as a watcher, a loop or a batch, the agent estimates the calls and runs:

node scripts/budget.mjs --calls 288 --input 3000 --output 300 \
  --model deepseek/deepseek-v4.1-flash --model anthropic/claude-sonnet-5.5

It reads the budget from GET /balance, prices each model from GET /models, and gives each one a verdict. With $5.00 to spend:

{
  "budget_usd": 5,
  "estimates": [
    {
      "model": "deepseek/deepseek-v4.1-flash",
      "total_usd": 0.082944,
      "budget_used_percent": 1.7,
      "verdict": "go"
    },
    {
      "model": "anthropic/claude-sonnet-5.5",
      "total_usd": 3.1104,
      "budget_used_percent": 62.2,
      "verdict": "tight"
    }
  ]
}

Verdict	When	The agent
`go`	Under half the budget	Starts, with `max_tokens` set on every call
`tight`	Between half and the whole budget	Tells you the numbers and offers a cheaper model or fewer calls
`over`	More than the budget	Doesn't start, and tells you

Half, because an agent's input grows with every turn and other calls share the same budget. With no flags, budget.mjs prints the budget alone: what new calls can spend, what running calls hold, the key's own limit and what expires soonest.

Models

node scripts/models.mjs --top 10                 # cheapest first
node scripts/models.mjs --reads image --top 10   # models that read images
node scripts/models.mjs --search sonnet          # find an id
node scripts/models.mjs --images                 # image models, with the fields each takes

The skill picks a cheap model for routine steps, like watching and summaries, and a strong one where the answer moves money. See Models.

Images

node scripts/vision.mjs --model <id that reads image> --image chart.png --prompt "Describe the trend."
node scripts/upload.mjs chart.png
node scripts/image.mjs --model <image model id> --prompt "..." --quality low

vision.mjs sends a small image in the call and uploads a larger one first. --image also takes an https link.
upload.mjs prints a link that works for 24 hours.
image.mjs prints a link to each image, kept 7 days, and the cost. Set --quality and --resolution where the model lists them: left out, the call reserves the largest the model offers.

Recap cards use only numbers you or the trading tools gave. The agent makes the card, reads it back with a model that reads images, and makes it again until every number matches. The skill never posts anywhere. See Images.

Rules it keeps

It never prints, logs or shares the key.
It asks before spending more than you set out to spend.
It never recommends or ranks a coin or token.
On Binance Agent OS, it reads the cockpit's status at each session start and never works around your trade limits. See The cockpit.

The source is on GitHub, under the MIT license.

InstallCopy Anchor Link

What to askCopy Anchor Link

Budget checkCopy Anchor Link

ModelsCopy Anchor Link

ImagesCopy Anchor Link