Inferable – AI Workflows with Human-in-the-Loop

TL;DR

TL;DR: Inferable is an open-source framework for building durable AI workflows with human approval gates, structured outputs, and versioned execution — self-hosted in your own infrastructure.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: inferable.ai — verified accessible
• Source repository: github.com/inferablehq/inferable — README read in full
• License: MIT (verified via GitHub API license.spdx_id)
• HN launch thread: news.ycombinator.com/item?id=42772862
• Source last checked: 2026-07-01 (commit main)

What Is Inferable?

Inferable is an open-source managed durable execution runtime for creating AI workflows with humans in the loop. The project positions itself as infrastructure for teams that want the power of LLM-driven automation but need guardrails — approval gates, versioned schemas, and traceable execution — without handing control entirely to the model.

From the README:

Create structured outputs from any LLM, ask humans for approval via Slack or Email, with versioned, long-running workflows for backwards compatibility.

Key capabilities confirmed from source:

• Self-hosted execution — Workflows run in your own infrastructure (on-prem, private VPC, behind a firewall). No inbound ports need to be opened; Inferable uses long-polling to connect outbound to the control plane.
• Human-in-the-Loop — Workflows can halt and wait for human approval via email or Slack before proceeding. Approval context is preserved across the interruption.
• Versioned workflows — Each workflow can have multiple versions with different input schemas. In-flight executions maintain affinity to the version they started with, enabling gradual rollouts.
• Structured outputs — LLM responses are parsed and validated against a Zod schema automatically, with built-in retries on parse failure.
• SDKs — Official Node.js/TypeScript (npm: inferable) and Go (pkg.go.dev/github.com/inferablehq/inferable/sdk-go) SDKs.
• Observability — Timeline view for workflow execution traces, plus plug-in points for external observability tools.
• Memoized results — Distributed caching of expensive side-effect results across workflow runs.

Setup Workflow

Prerequisites

• Node.js 18+ or Go 1.21+
• An Inferable account (free tier available) or a self-hosted control plane

Step 1: Install the SDK

Node.js / TypeScript:

npm install inferable

Go:

go get github.com/inferablehq/inferable/sdk-go

Step 2: Initialize a workflow

import { Inferable } from 'inferable';

const inferable = new Inferable({
  apiKey: process.env.INFERABLE_API_KEY,
});

const workflow = inferable.workflows.create({
  name: 'ticket-classifier',
  inputSchema: z.object({
    ticketText: z.string(),
  }),
});

workflow.version(1).define(async (ctx, input) => {
  const { ticketType } = ctx.llm.structured({
    input: `Classify this support ticket: ${input.ticketText}`,
    schema: z.object({
      ticketType: z.enum(['data-deletion', 'refund', 'technical', 'other']),
    }),
  });

  return { ticketType };
});

Step 3: Add a human approval gate

deleteUserWorkflow.version(1).define(async (ctx, input) => {
  if (!ctx.approved) {
    return Interrupt.approval({
      message: `Delete user ${input.userId}?`,
      destination: { type: 'email', email: '[email protected]' },
    });
  }

  await db.customers.delete({ userId: input.userId });
});

Step 4: Run and monitor

Workflow executions appear in the Inferable dashboard as a timeline, showing each step, LLM call, and approval event. Long-running workflows resume from their last checkpoint after a process restart.

Deeper Analysis

What makes it different from plain LangChain or LlamaIndex agents?

Most agent frameworks give you a prompt template and a tool-calling loop. Inferable’s value proposition is durable execution with explicit versioning and human oversight. Where a LangChain agent might silently retry or hallucinate its way past an error, Inferable:

• Halts on ambiguous decisions and surfaces them to a human
• Versions the workflow definition so schema changes do not break in-flight runs
• Provides a structured audit trail of every decision point

Infrastructure model

Inferable is unusual in that the control plane can be cloud-hosted or self-hosted, while the worker runs inside your network. This means your data never leaves your infrastructure for LLM calls — the worker polls outbound to the control plane, so no firewall holes are required.

For teams already running LangChain or similar: Inferable is not a replacement for the LLM abstraction layer, but rather a layer above it that adds workflow governance.

Practical Evaluation Checklist

• [ ] Installed Node.js SDK successfully (npm install inferable)
• [ ] Created a workflow with an input schema
• [ ] Verified structured output parsing with a Zod schema
• [ ] Tested approval interrupt with a mock email destination
• [ ] Confirmed long-poll connection works from a private network
• [ ] Checked timeline view in dashboard for execution trace

Security Notes

• Workflows execute entirely in your own environment — LLM calls are made from your infrastructure, not from Inferable’s servers.
• API key authentication is required for the control plane.
• No inbound ports are required; workers connect outbound via long-polling.
• For air-gapped environments, self-host the control plane and configure the worker to point to your internal endpoint.

FAQ

Q: Does Inferable host the LLM models? A: No. Inferable acts as the orchestration and governance layer. You bring your own LLM (OpenAI, Anthropic, local Ollama, etc.) — Inferable handles the workflow execution, not the model hosting.

Q: Can it run without internet (air-gapped)? A: Yes, if you self-host the control plane. Workers in isolated environments connect outbound to your self-hosted control plane via long-polling.

Q: What happens if a workflow crashes mid-execution? A: Workflows are durable. If a worker process crashes, the workflow resumes from its last checkpoint when a worker reconnects. No explicit resume step is needed.

Q: Is there a free tier? A: The README notes a free tier exists. For current pricing specifics, consult the Inferable pricing page.

Conclusion

Inferable fills a specific gap in the AI agent tooling landscape: teams that want the flexibility of LLM-driven workflows but need governance, versioning, and human oversight built in rather than bolted on. The self-hosted execution model is particularly compelling for regulated industries or teams with strict data residency requirements.

If you are building internal AI tools today and currently stitching together LangChain agents with ad-hoc retry logic and manual approval screenshots in Slack, Inferable is worth evaluating as a structured alternative.