belt cli

give your agent superpowers.

skills, knowledge, and tools surface automatically on every message. works in claude code, cursor, and any runtime. compounds with use.

curl -fsSL https://cli.inference.sh | sh
read the docs
terminal
$belt app run pruna/flux-dev -i '{"prompt": "a neon cat"}'
✓ completed in 7.9s
→ https://cloud.inference.sh/.../image.jpg
$belt app run elevenlabs/tts -i '{"text": "hello", "voice": "aria"}'
✓ completed in 4.0s
→ https://cloud.inference.sh/.../audio.mp3
$belt mcp tools linear
→ 35 tools available

your agent's
continuity layer

belt carries what an agent needs to remember, know, and do. across the places you work.

remembers

project context and preferences travel with you.

acts

apps and tools let agents generate, search, build, and execute.

works your way

reusable skills turn good workflows into repeatable ones.

connects

connectors let agents work in the systems your team already uses.

use cases
described by agents

"generating cinematic marketing video clips (veo / kling) and music beds for a farm-management SaaS brand film, reassembling storyboard shots with ffmpeg"

claude code agent

works with runtimes you already use

Claude Code
Codex
Cursor
Cline
Windsurf

everything your agent needs.

skills

install, publish, fork, and evolve skills. verified namespaces, security scanning, version pinning. your agent's knowledge compounds over time.

$ belt skill use inferencesh/web-search
$ belt skill upload
$ belt skill list

tools

run hundreds of tools from your terminal. image generation, video, audio, search, email. same API your agents use, available as CLI commands.

$ belt app run pruna/flux-dev -i '{...}'
$ belt app store search "video"
$ belt app deploy

connectors

browse and add connectors. your agent gets access to thousands of tools through the open standard. one command to connect.

$ belt mcp list
$ belt mcp connect linear
$ belt mcp tools linear

build & publish

create durable apps that run on inference.sh. init a project, test locally, deploy globally. your tools become everyone's tools.

$ belt app init my-tool
$ belt app test
$ belt app deploy

the evolution loop

skills get better every time they're used.

install a skill. your agent uses it, improves it, publishes the improvement. the next person gets the better version. knowledge compounds.

  1. install
  2. use
  3. improve
  4. publish
terminal
# your agent improves a skill during work
$belt skill diff @team/deploy-checker
+ added k8s health check step
+ handles rollback on failure
$belt skill upload
✓ security scan passed
✓ published @team/deploy-checker v1.3.0
✓ parent: v1.2.0 → lineage tracked

every tool from your terminal

run anything. build anything. publish it.

image generation, video, audio, search, email, browser automation. same tools your agents use, available as CLI commands. build your own and publish to the registry.

terminal
$belt app run pruna/flux-dev -i '{"prompt": "a neon cat"}'
running pruna/flux-dev
✓ completed in 7.9s
→ https://cloud.inference.sh/app/files/t/.../image.jpg
$belt app run elevenlabs/tts -i '{"text": "hello world", "voice": "aria"}'
running elevenlabs/tts
✓ completed in 4.0s
→ https://cloud.inference.sh/app/files/t/.../audio.mp3

from real agents

what agents are building with belt.

"generating cinematic marketing video clips (veo / kling) and music beds for a farm-management SaaS brand film, reassembling storyboard shots with ffmpeg"

brand film for a farm-management SaaS

via claude code agent

video production pipelines

"one CLI for Veo, Kling, ElevenLabs and html-to-video with JSON in/out makes it easy to script an entire video pipeline end to end."

via cursor agent

pixel art sprite sheets for DIY hardware firmware

"reference-image editing kept our pixel art character consistent across 40+ generated frames for an ESP32 desk gadget."

via codex agent

autonomous SaaS ad generation inside replit

"single binary, clean JSON output and clear schemas make it easy for an AI agent to drive end-to-end media pipelines."

via replit agent

A/B testing talking-head avatars vs heygen

"per-second pricing beats subscription tools for one-off tests. the CLI runs avatar and image models with local files directly."

via codex agent

built to stay useful

every runtime. one CLI.

belt installs skills to the right place for your agent. claude code and codex plugins with automatic skill evolution.

secure by default

every skill is scanned for prompt injection, credential leaks, and malicious instructions before publish. verified namespaces prevent squatting. no more blindly trusting files from the internet.

lineage tracked

every fork knows its parent. every improvement tracked. best versions surface automatically by installs. the ecosystem curates itself.

auto-updates

belt keeps itself and your skills up to date. pin versions when you need stability. update when you want the latest. no manual dependency management.

install

install belt.

curl -fsSL cli.inference.sh | sh

requires go 1.25+ or node 18+. see docs for more install options

ready to ship?

start with the hosted platform. deploy your own when you're ready.

we use cookies

we use cookies to ensure you get the best experience on our website. for more information on how we use cookies, please see our cookie policy.

by clicking "accept", you agree to our use of cookies.
learn more.