An AI workspace from Inference Shell Inc. in the United States. It lets you call AI models across image, video, audio, text, search, and 3D through a single API, regardless of which provider built them. On top of that, it provides a runtime for agents with state persistence, automatic retries, and human-in-the-loop approval steps. Because it absorbs the surrounding work — model selection, billing, and infrastructure — developers can concentrate on the logic of their agents and workflows. It supports models from major providers such as Anthropic, OpenAI, Google, xAI, and ElevenLabs, and pricing is based primarily on pay-per-run usage.
Key Features
- One API across many providers’ models: Through more than 100 pre-built apps called Tools, you can run image generation, video, audio, text, search, and 3D tasks from a unified API. There is no need to hold separate contracts and billing per provider
- A runtime for agents: A durable execution environment that can resume after a failure, with state retention, automatic retries, and human-in-the-loop flows built in. It suits long-running jobs and business processes that require a person to confirm a step
- Skills — a registry of reusable knowledge: Prompts and instructions can be registered as versioned “skills” and reused across multiple runtimes
- UI components: React components for chat screens and generative interfaces are provided, so an AI app can take shape without building a front end from scratch
- Commons and the Belt CLI: Commons is the shared memory and automation layer for a team, while the Belt CLI is a command-line tool that unifies skills, tools, connectors, and deployment. Python and JavaScript SDKs are also available
- External connectors and MCP support: Connectors for Slack, Discord, X, and Google Cloud are available, and the platform can also be connected as an MCP (Model Context Protocol) server
Pricing
| Plan | Monthly | Annual | What’s included |
|---|---|---|---|
| Starter | $0 | $0 | 1 concurrent task, 20 requests/min, 7-day retention, 1 seat, 100MB storage |
| Pro | $20 | $240 (about $16/mo) | $20 in credits, 5 concurrent tasks, 100 requests/min, 30-day retention, 1 seat, 5GB |
| Team | $200 | $2,400 (about $160/mo) | $200 in credits, 20 concurrent tasks, 500 requests/min, 90-day retention, 5 seats, 51GB |
| Enterprise | Contact sales | Contact sales | Custom arrangements for large organizations |
Separately from the credits included in each plan, running the models themselves is billed by usage. Rates are set per model and category — image generation from $0.001 per image, video by the second, chat models by token, and so on.
Pricing is as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- Contracts, API keys, and invoices for multiple AI providers can be consolidated into one
- Swapping models rarely requires large changes on the calling side, which helps avoid lock-in to a single vendor
- Durable agent execution, retries, and approval flows do not have to be built in-house
- A free Starter plan lets you test an implementation before spending anything
- The CLI, SDKs, and UI components together carry a project from prototype to production
⚠️ Cons
- With usage-based billing, costs are hard to predict for expensive workloads such as video generation
- Concurrency, request limits, and retention periods are fixed per plan, so heavy workloads assume a higher tier
- With many building blocks (Tools / Skills / Agents / UI / Commons / Belt), it takes time to grasp the whole picture
- If you want a given model’s newest features immediately, calling the provider’s own API directly can be faster
- Japanese-language documentation and case studies are still limited
Comparison with Similar Services
| Criteria | inference.sh | OpenRouter | Replicate | LangGraph Platform |
|---|---|---|---|---|
| Main role | Cross-provider API plus agent runtime | LLM routing and unified API | Hosting and running ML models | Platform for building and operating agents |
| Modalities | Image, video, audio, text, search, 3D | Mainly text (some multimodal) | Broad: image, video, audio, text | Depends on the framework |
| Agent execution | Durable execution, retries, approval flows built in | Not covered | Not covered (model execution focused) | Core capability |
| Pricing model | Subscription plus usage | Usage-based | Usage-based (by execution seconds) | Usage-based plus plans |
| Free tier | Yes (Starter) | Yes (some models) | Yes (trial credits) | Yes |
Who Is It For
- Developers juggling several AI providers whose API keys and invoices have become hard to manage
- Teams who need long-running AI jobs to survive interruptions, including retries and state retention
- Anyone building workflows with a human approval step, such as reviewing a draft or checking something before publication
- Anyone building an app that spans image, video, and audio without wiring up each API separately
- Anyone prototyping AI features on a stack that is not tied to one vendor
Summary
inference.sh bundles two things into one service: a unified layer that puts many providers’ models behind a single API, and a runtime that keeps agents running without falling over. If all you want is to call a model, simpler options exist, but when durable execution, retries, and human approval steps are part of the job, it becomes a real candidate. Start small on the free Starter plan, then consider Pro or Team once your usage becomes clear.