Inferable

What is Inferable?

Inferable is a managed durable-execution runtime for building reliable AI workflows and agents in code. Developers define versioned workflows in their own codebase and run them on their own compute, including private networks. A control plane preserves execution state, coordinates work, and connects to the runtime through outbound long polling, avoiding the need to expose inbound ports.

The platform provides structured LLM outputs with validation, agent tool use, memoized results, notifications, timeline observability, retries, long-running execution, and human approvals through email or Slack. Workflows can pause while waiting for a person or agent and later resume with context intact. The repository contains the open-source control plane, management application, CLI, and SDKs for multiple languages.

How to Use Inferable

1. Create a cluster and install the SDK for the application's language.
2. Define a workflow name, input schema, execution identifier, and version in code.
3. Add structured model calls, agent steps, tools, side effects, and output schemas.
4. Insert approval or intervention points for sensitive actions.
5. Start the workflow listener in the application's own compute environment.
6. Trigger executions with unique idempotency identifiers and inspect the timeline for progress or failures.
7. Add a new version when changing schemas or logic while older executions finish on their original version.

Core Features

• Workflow as code: Defines multi-step AI orchestration inside the application's repository.
• Durable execution: Persists workflow state across long waits, network failures, and process interruptions.
• Human in the loop: Pauses for approval or input through email or Slack with context preserved.
• Structured outputs: Parses and validates model responses and retries failed executions.
• Versioned workflows: Maintains compatibility for in-progress executions during changes.
• Own-infrastructure execution: Runs functions behind firewalls or private networks with outbound connectivity.
• Agents and tools: Supports iterative reasoning engines that call registered application functions.
• Observability and memoization: Provides execution timelines and caches expensive or side-effecting results.

Use Cases

• Approval-sensitive automation: Require a person before deleting data, sending a decision, or taking another high-impact action.
• Long-running AI processes: Pause for external events and resume without losing workflow state.
• Structured document processing: Extract validated fields from content and retry invalid model output.
• Internal agents: Give an agent controlled tools inside a private application environment.
• Version-safe production workflows: Roll out new logic while old executions complete predictably.
• Distributed orchestration: Coordinate work across environments without exposing internal inbound endpoints.

Pricing

The Free tier is listed at $0 with two registered workflows, 1,000 workflow executions, bring-your-own models, and maximum concurrency of two. Pay as you go is listed at $10 per month with 20 workflows, 5,000 included executions, additional executions at $1 per 1,000, concurrency of ten, higher limits, and priority email support. Enterprise offers custom cloud deployments, data isolation, dedicated Slack support, and custom service levels. Model-provider costs may be separate when bringing your own models.

Frequently Asked Questions

Where does workflow code execute?

It runs in the user's own infrastructure, while the control plane coordinates execution and preserves state.

Can a workflow wait for human approval?

Yes. Inferable can pause, notify a reviewer through email or Slack, and resume with context preserved.

Is Inferable open source?

Yes. The repository includes the control plane, management application, CLI, and several SDKs under an MIT license.