Belt

belt cli

belt is the cloud platform cli for ai agents. single ~4mb binary, no runtime dependencies.

using a purpose-built cli means your agent operates through a constrained, typed interface instead of writing raw curl commands or sdk calls. every operation goes through schema validation — fewer tokens, fewer errors, and no credential leakage.

install

first check if belt is already installed:

which belt && belt --version

if already installed, skip to authenticate.

package managers (recommended — verified through each registry's trust chain):

brew install inference-sh/tap/belt            # macos / linux (homebrew tap, signed)
scoop bucket add belt https://github.com/belt-sh/scoop-belt && scoop install belt  # windows
npm install -g @inferencesh/belt              # node.js (global install, pinned in package.json)

manual install (full control — download, verify, then run):

curl -fsSL https://cli.inference.sh -o /tmp/belt-install.sh

the installer is a short, readable shell script. it detects your os and architecture, downloads the matching binary from dist.inference.sh, verifies the binary's sha-256 checksum against the published manifest, and places it in your path. no elevated permissions required. the installer source is publicly readable — review it before running:

cat /tmp/belt-install.sh   # review the script
sh /tmp/belt-install.sh    # run after review

authenticate

belt login
belt me

set up agent integration

belt plugin init claude     # claude code
belt plugin init codex      # openai codex
belt plugin init cursor     # cursor

quick start

belt suggest "what tool should i use"  # unified search across apps, skills, knowledge
belt app store                         # browse ai apps
belt app store --category video        # filter by category
belt app get <namespace/name>          # see schema, pricing, functions
belt app sample <namespace/name>       # generate sample input json
belt app run <namespace/name> --input input.json  # run an app
belt balance                           # check credits

common workflows

image and video generation:

belt app get bytedance/seedance-2-0           # check schema — file fields accept local paths
belt app sample bytedance/seedance-2-0 --save input.json
# edit input.json, then:
belt app run bytedance/seedance-2-0 --input input.json

file fields (type: file in schema) accept local paths directly — the cli auto-uploads them:

belt app run bytedance/seedance-2-0 --input '{"image": "./photo.jpg", "prompt": "make it cinematic"}'

check pricing before running:

belt app pricing <namespace/name>             # see cost formula
belt app pricing <namespace/name> --json      # machine-readable

multi-function apps (e.g. apps with list_voices, list_resources, etc.):

belt app get heygen/avatar-video              # shows all functions with schemas
belt app sample heygen/avatar-video -f list_resources   # sample for a specific function
belt app run heygen/avatar-video -f list_resources --input '{}'

knowledge and skills:

belt knowledge list --json                    # your knowledge entries
belt knowledge search "react patterns"        # semantic search
belt skill list                               # your skills
belt skill store search "video"               # find skills in the store
belt skill use <namespace/name>               # load a skill on-demand

mcp connectors:

belt mcp list                                 # browse available connectors
belt mcp connect slack                        # connect one
belt mcp run slack send_message --input '{"channel": "#general", "text": "hello"}'

machine-readable output:

all list commands support --json for structured output:

belt app list --json
belt app store --json
belt task list --json
belt knowledge list --json
belt skill list --json
belt mcp list --json
belt secrets list --json
belt me --json
belt balance --json

tips

• belt app sample generates ready-to-edit input json from the app schema
• file fields show ./your-file.jpg in samples — just replace with your actual file path
• belt suggest searches apps, skills, and knowledge in one call
• belt task cost <task-id> shows actual cost after a run
• belt app run --no-wait submits without blocking, belt task get <id> to check later
• use --session new for stateful apps that keep gpu warm between calls

disable hooks for a project

# .beltsh/config.json
{"hooks_disabled": true}

or set BELT_NO_HOOKS=1 in your environment.

links

belt.sh · docs · trust · source