how is inference shell different from Replicate?

Replicate has AI models. inference shell has AI models plus video rendering, email, search, project management, and MCP servers. all composable. plus BYOK: bring your own keys to route through Fal, Google, or your own GPUs.

Model Context Protocol, an open standard for connecting tools to software. inference shell is an MCP server: every app on it runs from Claude Code, Cursor, or any MCP client, and you can add MCP servers for other services.

can I use apps without building agents?

yes. call any app with a single HTTP request or compose them into workflows with flows. no agents required.

apps

any ai model. one api call.

399 apps across image, video, audio, chat, 3d and search. one key, pay per run, zero vendor lock-in.

start free read the api docs

browse by category

image

94 apps

video

99 apps

audio

45 apps

2 apps

text

27 apps

chat

68 apps

decision

18 apps

social

25 apps

utilities

1 apps

other

20 apps

featured

Seedance 2.5

Professional multimodal video generation from text, images, video, and audio references using ByteDance's Seedance 2.5 model via BytePlus ARK API. Supports up to 4K (10-bit color), durations up to 30s, MOV output, and multimodal reference-to-video with synchronized audio.

P-Video-Avatar

Generate talking head videos from a portrait image with text or audio-driven speech

Veo 3.1 Fast

Veo 3.1 Fast via Vertex AI - Generate videos from text prompts or images with optional audio

ByteDance USO

A unified image editor that allows users to generate images by combining any subject with any style efficiently, preserving identity and consistency.

most used

H3 Max Styles

falai/minimax-h3-max-styles

stylized 768p video with audio from a prompt or first-frame image using fal's minimax h3 max style pipelines. vhs, hand-drawn, 16-bit pixel, retro 70s toon, low-poly.

H3 Max Insert Video

falai/minimax-h3-max-insert-video

insert a newly generated scene into an existing video with minimax h3 max via fal. pick where the scene starts and where the original resumes; guide it with a prompt and references.

H3 Max 3D to Video

falai/minimax-h3-max-3d-to-video

turn a blender or 3d previs render into photorealistic video with minimax h3 max via fal. layout, camera and timing follow the input; optional reference images set the look.

H3 Max Recast

falai/minimax-h3-max-recast

replace the people in a video with those in reference photos using minimax h3 max via fal. keeps motion, camera, cuts and audio. up to 4 people, 768p or 1080p.

H3 Max Relight

falai/minimax-h3-max-relight

relight an existing video from a lighting sphere reference image with minimax h3 max via fal. keeps subjects, motion, camera and audio; output 480p to 2k.

H3 Max Lip Sync

falai/minimax-h3-max-lip-sync

animate a portrait image to speak an audio track with minimax h3 max via fal. output matches the audio length (5s to 15 min), 480p to 2k, optional transcription guidance.

H3 Max Camera Controls

falai/minimax-h3-max-camera-controls

turn a single image into a video with keyframed camera orbit, elevation and distance using minimax h3 max via fal. the scene stays frozen; only the camera moves.

H3 Max Extend Video

falai/minimax-h3-max-extend-video

extend an existing video with new prompt-guided footage using minimax h3 max via fal. 1-15s of continuation, 480p to 2k, optional audio references.

H3 Max Turbo Extend Video

falai/minimax-h3-max-turbo-extend-video

extend an existing video with new prompt-guided footage using minimax h3 max turbo via fal. 1-15s of continuation, 480p to 2k, optional audio references.

H3 Max Turbo

falai/minimax-h3-max-turbo

minimax h3 max turbo via fal, the faster, cheaper h3 max variant. text-to-video and image-to-video with first/last frame, 480p/768p/1080p, 1-15s, prompt expansion, audio pinning.

FLUX 3 Max

falai/flux-3-max

flux 3 max, post-trained flux 3 video by fal and black forest labs. text-to-video, image-to-video with end frame, keyframes-to-video. 720p or 1080p, 5-20s, native audio.

DeepSeek V4.1 Flash 1017

melious/deepseek-v4-1-flash-1017

deepseek v4.1 flash 1017 — the october 2026 refresh of the 552b multimodal moe with 1m context, image input and controllable reasoning. via the melious api.

Google Analytics

google/analytics

read google analytics 4 properties and run traffic, acquisition and realtime reports

Google Search Console

google/search-console

read search console performance, url index status, and sitemaps for your verified sites

Fish Audio Drama 3

fish-audio/drama-3

fish audio drama 3 text to speech. direct the delivery in plain language, add laughs and sighs, clone a voice from a clip, or voice a dialogue. 80+ languages.

Gemini 4 Argon

google/gemini-4-argon

gemini 4 argon via vertex ai — google's flagship model for deep, multi-step work in coding, finance, law, security and science. 1m token context, thinking levels low to high.

GPT-6 Sol

openai/gpt-6-sol

gpt-6 sol — openai's gpt-6 mid-tier model, superseded by gpt-6.1 sol. 1m context, 128k output, reasoning effort none to max, vision, file input and tool use via the responses api. direct api.

GPT-6 Luna

openai/gpt-6-luna

gpt-6 luna — openai's most efficient model for focused, high-volume tasks. 1m context, 128k output, reasoning effort none to max, vision, file input and tool use via the responses api. direct api.

GPT-6.1 Sol

openai/gpt-6-1-sol

gpt-6.1 sol — openai's near-astra flagship at a fraction of the cost. 1m context, 128k output, reasoning effort low to max, vision, file input and tool use via the responses api. direct api.

Claude Fable 5.1

anthropic/claude-fable-5-1

claude fable 5.1 — anthropic's most capable model for demanding reasoning and long-horizon agents. 1m context, 128k output, always-on thinking with effort, vision, tool use. direct api.

Claude Haiku 5.5

anthropic/claude-haiku-5-5

claude haiku 5.5 — anthropic's fastest, lowest-cost claude for high-volume and latency-sensitive work. 1m context, 128k output, adaptive thinking, vision, tool use. direct api.

Claude Sonnet 5.5

anthropic/claude-sonnet-5-5

claude sonnet 5.5 — anthropic's fastest frontier sonnet for coding, agents and everyday work. 1m context, 128k output, adaptive thinking with effort, vision, tool use. direct api.

Claude Opus 5.5

anthropic/claude-opus-5-5

claude opus 5.5 — anthropic's default opus for long-running agentic coding and knowledge work. 1m context, 128k output, adaptive thinking with effort, vision, tool use. direct api.

Muse Wonder 1.0

meta/muse-wonder-1-0

muse wonder 1.0 — meta's most capable model for coding and long-horizon agents. 1m context, reasoning minimal to max, image, video, audio and pdf input, tools. direct api.

Decision 3.0 d3 27B

vllm-sr/d3

the largest decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

Decision 3.0 d3-flash 9B

vllm-sr/d3-flash

the 9b decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

Decision 3.0 d3-mini 4B

vllm-sr/d3-mini

the 4b decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

Decision 3.0 d3-nano 2B

vllm-sr/d3-nano

the 2b decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

Decision 3.0 d3-lite 0.8B

vllm-sr/d3-lite

the 0.8b decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

Decision 3.0 d3-edge 0.6B

vllm-sr/d3-edge

the smallest and fastest decision 3.0 model from vllm semantic router. send text, images and videos with typed questions (choice, score, yes/no); get a probability for every answer.

HTTP Request

infsh/http

make an http request to a public url and get the status, headers and body back. json, form, text and multipart file uploads; internal addresses are refused.

Gemini 3.1 Pro Preview

openrouter/gemini-3-1-pro-preview

google's frontier reasoning model, with stronger coding and agentic reliability. 1m context; text, image, file, audio and video input.

OpenAI Decisions

openai/decisions

openai's decisions api on gpt-6 luna. send text and images with typed questions (choice, score, yes/no); get a probability for every answer in near real time, with no text generation. direct api.

d1-omni-600M

liquid/d1-omni-600m

liquid ai's 587m decision model for text, images and speech. send a state with images or an audio clip and typed questions (choice, score, yes/no); get a probability for every answer. research release

d1-3B

liquid/d1-3b

liquid ai's open 3b decision model for text and images. send a state and up to 8 images with typed questions (choice, score, yes/no); get a probability for every answer in one pass, no generation.

Perplexity Decider v1.1 27B

perplexity/decider-v1-1-27b

perplexity's open decision model for text and images. send a state, up to 4 images and typed questions (choice, score, yes/no); get a calibrated probability for every answer, with no text generation.

Decision 2.0 Vega 27B

vllm-sr/decision-2-0-vega-27b

the largest decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one pass. 29.37b parameters, 32k-token input.

JEV-27B-VL

autotrust/jev-27b-vl

autotrust's open decision model with vision. send text and up to 8 images with typed questions (choice, score, yes/no); get a calibrated probability for every answer, with no text generation.

Gemini Nano Banana 2.1

google/gemini-nano-banana-2-1

gemini nano banana 2.1 via vertex ai — google's newest image model. better design, mask editing and subject consistency than nano banana 2, and faster. up to 4k, web search grounding.

Decision 2.0 Lux 9B

vllm-sr/decision-2-0-lux-9b

the 9b decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one forward pass, with no text generation. 7.94b parameters, 16,384-token input, about 18.4 ms per question.

Decision 2.0 Nox 4B

vllm-sr/decision-2-0-nox-4b

the 4b decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one forward pass, with no text generation. 4.21b parameters, 16,384-token input, about 12.9 ms per question.

Decision 2.0 Sol 2B

vllm-sr/decision-2-0-sol-2b

the 2b decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one forward pass, with no text generation. 1.88b parameters, 16,384-token input, about 7.2 ms per question.

Decision 2.0 Eos 0.8B

vllm-sr/decision-2-0-eos-0-8b

a small, fast decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one forward pass, with no text generation. 0.75b parameters, 16,384-token input, about 6.0 ms per question.

Decision 2.0 Kai 0.6B

vllm-sr/decision-2-0-kai-0-6b

the smallest and fastest decision 2.0 model from vllm semantic router. send a state and typed questions (choice, score, yes/no); get a probability for every answer in one forward pass, with no text generation. 0.60b parameters, 8,192-token input, about 4.9 ms per question.

Eleven Flash v2.5

elevenlabs/eleven-flash-v2-5

elevenlabs eleven flash v2.5 text to speech. the fastest and lowest-cost elevenlabs model, for long texts and high volume.

Eleven Multilingual v2

elevenlabs/eleven-multilingual-v2

elevenlabs eleven multilingual v2 text to speech. stable, high-quality multilingual speech for long-form content.

Eleven v3

elevenlabs/eleven-v3

elevenlabs eleven v3 text to speech. expressive speech in 70+ languages with audio tags for emotion and delivery.

Eleven v4 Turbo

elevenlabs/eleven-v4-turbo

elevenlabs eleven v4 turbo text to speech. the low-latency v4 model; send a text for an audio file, or stream text over a socket and hear it as it is written.

Eleven v4

elevenlabs/eleven-v4

elevenlabs eleven v4 text to speech. the newest and most expressive model, with 90+ languages and audio tags; send a text for an audio file, or stream text over a socket and hear it as it is written.

GPT Transcribe

openai/gpt-transcribe

transcribe speech with openai. send a recording and get its transcript from gpt transcribe, or stream a microphone over a socket and read the transcript as it is spoken with gpt live transcribe.

Grok Speech to Text

xai/grok-stt

transcribe speech with xai's grok speech to text. send a recording and get the transcript with word timings and speakers, or stream a microphone over a socket and read the transcript as it is spoken.

Gemini 3.7 Flash

google/gemini-3-7-flash

gemini 3.7 flash via vertex ai — improved software engineering, web development and agentic workflows over 3.6 flash. 1m token context window.

Gemini 3.8 Flash

google/gemini-3-8-flash

gemini 3.8 flash via vertex ai — google's most capable flash model, built for long-horizon software engineering, autonomous agents and complex enterprise workflows. 1m token context window.

DeepSeek V4.1 Flash

melious/deepseek-v4-1-flash

deepseek v4.1 flash — 552b multimodal moe with 1m context, image input and controllable reasoning. frontier agentic and coding performance. via the melious api.

MiniMax M3

melious/minimax-m3

minimax m3 — 428b multimodal moe with 1m context and image input for long-horizon agentic, coding and cowork tasks. via the melious api.

GLM 5.3

melious/glm-5-3

glm 5.3 — zai's 744b moe post-trained for complex coding and long-horizon agentic tasks. 1m context, thinking on by default. via the melious api.

Kimi K3

melious/kimi-k3

kimi k3 — moonshot's 2.8t open-weight multimodal agentic model with native vision and 1m context for long-horizon coding and reasoning. via the melious api.

Qwen 3.6 27B

melious/qwen3-6-27b

qwen 3.6 27b — dense 27b vision-language model with 262k context and thinking mode. strong coding and multimodal reasoning, apache 2.0. via the melious api.

GLM 5.3 Flash

melious/glm-5-3-flash

glm 5.3 flash — zai's 320b moe (18b active), the first natively multimodal glm-5 model. 1m context, image input, low cost. via the melious api.

GLM 5.2

melious/glm-5-2

glm 5.2 — zai's 744b moe with 1m context and hybrid reasoning for coding and agentic tasks. mit license. via the melious api.

not enough? create new apps fast. templates + coding agents make it insanely extensible.

create your own apps

start from templates. add code, packages, docs. deploy in minutes.

$ infsh app init

my-app/

inference.py

requirements.txt

$ infsh app deploy

schemas become tool parameters automatically. your app shows up in the grid and can be used by agents and workflows.

read the app creation guide

create workflows

build a graph of apps. deploy as a single callable app.

drag and drop to build the graph. map io to connect steps. deploy as an app.

read the workflow guide

three sources of apps

built-in. connected. composed.

every app funnels into the same interface. where it comes from doesn't matter. same shape, same API.

built-in

hundreds of apps

image generation, video, audio, chat, search, and more. run any model with a single API call. pay per run, no subscriptions.

connected

MCP servers

browse and add MCP servers. linear, notion, slack, github. your agent gets thousands of tools through the open standard.

composed

flows become apps

chain any apps into a flow. the result is callable like any other app. AI models, non-AI apps, MCP servers, other flows. all composable.

also available as an MCP server

use from claude code, cursor, or any MCP client

add one line to your MCP config. every app on inference shell becomes available to your agent.

mcp config

{
  "mcpServers": {
    "inference": {
      "url": "https://api.inference.sh/mcp"
    }
  }
}

one api call

run any app in three lines.

terminal

import Inference from '@inference/sdk'

const client = new Inference()

const result = await client.run('pruna/flux-dev', {

prompt: 'a marble workbench with glowing tools'

})

console.log(result.image)

// → https://cloud.inference.sh/...

output

Generated: a marble workbench with glowing tools

pruna/flux-dev · 7.9s · $0.005

sdks: javascript · python · go · cli

built to stay useful

one api. your keys. durable execution.

one api

same interface for every app. built-in or connected, same shape: input in, output out. no per-provider SDKs.

byok

bring your own keys. route model runs through fal, google, or your own GPUs. you're not locked in.

durable

every app run retries on failure, persists state, and tracks execution. built in, not bolted on.

belt

run any app from your terminal. belt app run pruna/flux-dev -i '{...}'. one command, same api.

curl -fsSL https://cli.inference.sh | sh

inference shell vs Replicate inference shell vs Fal inference shell vs calling APIs directly

faq

frequently asked questions

can't find what you're looking for? we're here to help.

an ai runtime where everything compounds. run any model, compose agents, stack knowledge. also includes a skill registry, MCP servers, 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.

learn

build

from the blog