how is inference.sh different from Replicate?

Replicate has ai models. inference.sh has ai models plus video rendering, email, search, project management, and mcp servers, all composable through one api. plus byok (bring your own keys) to route through Fal, Google, or your own GPUs.

can I use inference.sh without building agents?

yes. call any tool with a single HTTP request or compose them into workflows. no agents required. agents, skills, and teams are separate products on the same platform.

bring your own keys. route model runs through Fal, Google, or your own GPUs. you're not locked in to any single provider.

one api for everything.

250+ serverless tools. ai models, video, audio, search, social, dev tools. call anything, compose everything.

get api key browse tools

terminal

$belt app run pruna/flux-dev

✓ completed in 7.9s

$belt app run bytedance/seedance-2

✓ completed in 42s

$belt app run elevenlabs/tts

✓ completed in 4.0s

$belt app run tavily/search-assistant

✓ completed in 1.2s

$belt app run x/post-create

✓ completed in 0.8s

$belt app run openai/gpt-image-2

✓ completed in 12s

$belt app run infsh/remotion-render

✓ completed in 8.4s

$belt app run openrouter/claude-opus-46

✓ completed in 3.1s

$belt app run infsh/agent-browser

✓ completed in 5.2s

$belt app run pruna/p-image-upscale

✓ completed in 2.1s

$belt app run google/veo-3

✓ completed in 38s

$belt app run alibaba/wan-2-7-r2v

✓ completed in 26s

$belt app run google/gemini-3-1-flash-image

✓ completed in 6.2s

$belt app run bytedance/seedream-4-5

✓ completed in 4.8s

$belt app run xai/grok-imagine-image

✓ completed in 8.1s

$belt app run pruna/qwen-image

✓ completed in 3.4s

$belt app run elevenlabs/sound-effects

✓ completed in 6.7s

$belt app run elevenlabs/voice-changer

✓ completed in 2.3s

$belt app run infsh/shell

✓ completed in 0.4s

$belt app run alibaba/happyhorse-1-0-t2v

✓ completed in 31s

$belt app run pruna/p-video-avatar

✓ completed in 18s

$belt app run bytedance/omnihuman-1-5

✓ completed in 22s

$belt app run tavily/extract

✓ completed in 1.8s

$belt app run x/user-get

✓ completed in 0.3s

$belt app run pruna/wan-t2v

✓ completed in 35s

$belt app run xai/grok-tts

✓ completed in 3.9s

$belt app run infsh/image-resize

✓ completed in 0.6s

$belt app run alibaba/qwen-image-2-pro

✓ completed in 5.1s

$belt app run bytedance/seedream-5-lite

✓ completed in 4.2s

$belt app run pruna/flux-klein-4b

✓ completed in 2.8s

the platform

tools

250+ built-in, connected, and composed. ai models, video rendering, email, search. one api.

explore

skills

the skill registry. versioned, secure, evolving. works in Claude Code, Cursor, and every runtime.

explore

agents

ship agents that run in production. durable execution, human-in-the-loop, 250+ tools.

explore

ui

ai-native react components. chat, generative ui, tool approvals. 30+ widgets.

explore

teams

coming soon

your team's ai workspace. shared memory, automations, self-hostable.

explore

belt cli

the cli that ties it all together. run tools, manage skills, connect mcp servers, deploy.

explore

why we built this

the pain points that made us build inference.sh.

the moat

"the agent framework is not the moat. prompt engineering is not the moat. the base LLM is not the moat."

"the specialized tools that encode domain knowledge - are the moat."

observability

"i spent 6 hours debugging a workflow that had zero error logs. when something breaks at 2 AM, i don't want to trace through 47 nodes."

"i want to see exactly what payload caused the issue."

durable execution

"i felt like a 'button person' in my IDE. the agent works in quanta - cut off by time every 2 minutes."

"long tasks require pipeline thinking, not chat sessions."

real-time streaming

"our multi-step agent produced great results but took 45+ seconds. users thought it crashed."

"if they see the internal monologue, they wait. if they see a spinner, they leave."

pay-per-execution

"spent 10 hours deploying agents on EC2. switched to serverless."

"why is this so hard?"

decision context

"systems record that a ticket was escalated, but not why it happened."

"without that reasoning, agents treat every edge case as a brand new problem."

frequently asked questions

one api for everything. 250+ serverless tools: ai models, dev tools, and integrations. run them directly, compose them into workflows, or use them to power agents. also includes a skill registry, react components, and a team workspace.

ready to ship?

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

start for free view on github

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.