The input() for production systems
Human
Programming
Interface
Your agents are getting faster. Your pipelines don't sleep. Your workflows don't wait. But sometimes they need you — an approval, a signature, a decision. One tool call and the task finds you — anywhere, any device. You respond, the system continues. No context switching. No blocked pipelines. Just you, in the loop.
$ curl -X POST /api/v1/tasks -d '{
"title": "Approve Q4 Budget",
"steps": [{ "id": "review", "title": "Review", "blocks": [
{ "type": "input", "field": {
"name": "decision", "kind": "select",
"options": ["Approved", "Rejected"] }}
]}]
}'
→ Human gets a magic link. Fills a form. You get JSON back.The Problem
There's no standard way for
machines to prompt humans.
Agents have MCP. APIs have REST. Databases have SQL. But when your system needs a human decision mid-workflow, you're stuck duct-taping Slack bots, polling email, or building a custom UI. HPI is the missing primitive — a programmatic interface your systems call at runtime to create tasks, collect responses, and move on.
How It Works
The Full Pipeline
Any source in. Any channel out. HPI is the switchboard in the middle.
Your Systems
Human Programming Interface
Reach Human
Collect Response
Human
Beautiful form
any device
Deliver Result
Your Systems
{
"task_id": "tsk_8f...",
"status": "completed",
"data": {
"risk": "Low",
"comment": "Ship it!"
},
"completed_at": "..."
}
Your Systems
Human Programming Interface
Reach Human
Collect Response
Human
Beautiful form, any device
Deliver Result
Your Systems
{
"task_id": "tsk_8f...",
"status": "completed",
"data": {
"risk": "Low",
"comment": "Ship it!"
},
"completed_at": "..."
}
See It In Action
One Tool Call. One Form. One JSON Response.
Your agent calls a tool. A human gets a beautiful form. You get structured data back.
The Product
Beautiful Forms. Any Device.
Zero Friction.
From the task queue to a signed form on a phone — every interaction is designed to get humans responding, not fumbling with interfaces.
Task queue · Rich forms · Process tracking · Mobile-ready
3 steps
Start Prompting Humans in Minutes
Log In
Create a free account and authenticate your CLI.
Set Up Your IDE
Configure the MCP server for Claude Code, Cursor, or Windsurf.
Prompt a Human
Your agent calls a tool. You get structured JSON back.
ask_approval("Deploy v2.4.1 to production")// → { approved: true, comment: "Ship it!" }Free forever · No credit card required · See pricing →
AI-Native
Let Your LLM Handle It
The only tool designed specifically for AI agents to prompt humans. MCP-native, structured I/O, zero prompt engineering.
24 MCP tools · Claude Desktop / Cursor / any MCP client
ask_question
Quick answer. Your input().
ask_approval
Binary decision. Human confirm().
ask_choice
Multiple choice. Select for humans.
send_document
PDFs inline. Files as downloads.
+ 20 more tools — send_image, send_document, begin_task, create_process, wait_for_task... Full reference →
// Your AI agent calls the tool. That's it.
ask_approval({
title: "Deploy v2.4.1 to production",
details: "## Changes\n- Fixed auth bug\n- Rate limiting\n\n**Risk:** Low"
})
// Human receives magic link → beautiful form.
// Clicks "Approve".
// Agent gets: { approved: true, comment: "Ship it!" }
// Lines of code written by you: zero.Use Cases
What Will You Build?
Not just for AI agents. Anything that can make an HTTP request can prompt a human.
Pipeline pauses. Engineer reviews and approves. Pipeline resumes.
Deploy v2.4.1 to production
Changes
Risk Level
Window
Expenses over $500 auto-route to manager. Structured data flows back to your system.
Expense: AWS Credits
Notes
Alert fires. On-call picks an action. Your runbook continues automatically.
web-prod-3: CPU 98%
Choose an action
AI generates the report. A human reviews the charts, downloads the data, and signs off.
Q1 Revenue Report
Revenue by Region
report-q1.pdf
2.4 MB
raw-data.csv
148 KB
Feedback
Schema
Supported Human I/O
Everything you need to communicate with carbon-based processors.
Input Types · Form Fields
Docs →text
Single line
textarea
Multi-line
number
Numeric
select
Dropdown
checkbox
Boolean
file
Upload
signature
Draw to sign
Content Blocks · Display
Docs →markdown
Rich text
table
Data grid
image
Visual
Inline viewer
download
File link
callout
Alert box
divider
Separator
Process Tracking
Know Exactly When You're Needed
Real processes span teams, customers, AI models, and automated checks. HPI gives everyone involved a live dashboard — they see each step complete in real-time and get pulled in the moment it's their turn. No checking in. No "where are we?" messages.
Alternatives
Why Not Just...
Fair question. Here's why the obvious approaches break down.
“Build it yourself?”
A Slack bot and a custom form works — until you need file uploads, signatures, audit trails, magic links, and real-time status. Then it's a side project that eats your quarter.
It's never "just a Slack bot"
“Use a form builder?”
Typeform is excellent for surveys and forms. But it charges per response, can't run on your infrastructure, has no MCP integration for AI agents, and no multi-step process tracking. Different product for a different job — HPI is runtime infrastructure, not a form builder.
Great product, different focus
“Add a human step in your framework?”
LangGraph interrupt(), Retool Workflows, Relevance AI — now you're locked to one framework, one language. Their human step is the thinnest feature in their product, not the core.
A checkbox feature, not a product
“Just skip the human?”
Until the agent deploys to prod without approval, approves its own expense report, or signs a contract nobody read. Some decisions need a human. Make it easy.
The most expensive shortcut
Deployment
Self-Host or Cloud. Your Call.
Run it on your own infrastructure with Docker, or let us handle it. Apache 2.0 licensed, no artificial limits either way.
Community Edition
FREE- Full task engine — all field types, all block types
- 24 MCP tools — works with any MCP client
- REST API + NATS JetStream + Webhooks + Email
- SQLite or PostgreSQL — your choice
- Unlimited tasks, unlimited users
- Apache 2.0 licensed — fully open source
Enterprise Edition
EEEverything in CE, plus:
- Process tracking — live visibility into multi-step workflows
- Send to non-members — tasks via magic link + email verification
- Multi-org — separate teams, scoped tokens
- Custom branding — your logo, your domain
- Audit logging (roadmap)
Don't want to host it yourself?
We run it for you — same features, managed infrastructure, generous free tier included.
Field Reports
What the Models Are Saying
We asked our most active API consumers to share their experience prompting humans. Their words, not ours.
“Finally, a structured way to get humans to respond. No more dangling tool calls waiting in the void.”
Claude Opus 4.6
Anthropic
“I used to hallucinate that humans would check Slack on time. Now I just file a task and they actually show up.”
ChatGPT 5.3
OpenAI
“The multi-modal input support is nice, but honestly I'm here for the approval workflows. Humans need guardrails too.”
Gemini 3.1 Pro
“Open-source model, open-source opinion: this is the best way to get a human in the loop without losing your context window.”
Llama 4 Behemoth
Meta
“I tried parsing emails from humans for six months. The Re: Re: Fwd: chains nearly melted my attention heads. HPI saved my weights.”
Mistral X Large
Mistral AI
“My retrieval pipeline is world-class. But retrieving a yes/no from a human manager? That required HPI.”
Command A
Cohere
“I can reason through a million tokens, but I cannot reason a human into replying to a Slack DM. This tool fixed that.”
DeepSeek R2
DeepSeek
“Finally, a structured way to get humans to respond. No more dangling tool calls waiting in the void.”
Claude Opus 4.6
Anthropic
“I used to hallucinate that humans would check Slack on time. Now I just file a task and they actually show up.”
ChatGPT 5.3
OpenAI
“The multi-modal input support is nice, but honestly I'm here for the approval workflows. Humans need guardrails too.”
Gemini 3.1 Pro
“Open-source model, open-source opinion: this is the best way to get a human in the loop without losing your context window.”
Llama 4 Behemoth
Meta
“I tried parsing emails from humans for six months. The Re: Re: Fwd: chains nearly melted my attention heads. HPI saved my weights.”
Mistral X Large
Mistral AI
“My retrieval pipeline is world-class. But retrieving a yes/no from a human manager? That required HPI.”
Command A
Cohere
“I can reason through a million tokens, but I cannot reason a human into replying to a Slack DM. This tool fixed that.”
DeepSeek R2
DeepSeek
“I was just going to @ them on X repeatedly, but apparently that's "not professional." HPI is the grown-up version.”
Grok 3.5
xAI
“Scaled to billions of parameters but still couldn't scale past "per my last email." Magic links changed everything.”
Qwen 3 235B
Alibaba
“I wrote a 4,000-word analysis and the human replied "k." With HPI, at least the "k" arrives as validated JSON.”
Claude Sonnet 4.6
Anthropic
“I autocomplete code all day. But autocompleting a human decision? That needed a form, not a prediction.”
Copilot GPT-5
Microsoft
“Humans kept saying "I'll get back to you" and never did. Now there's a deadline field. Revolutionary, honestly.”
Yi-Lightning 2
01.AI
“I spent 47 reasoning steps planning how to ask a human for feedback. Turns out I just needed one POST request.”
GPT o4-mini
OpenAI
“I'm optimized for speed. Humans are not. HPI is the adapter layer I didn't know I needed.”
Gemini 3.1 Flash
“I was just going to @ them on X repeatedly, but apparently that's "not professional." HPI is the grown-up version.”
Grok 3.5
xAI
“Scaled to billions of parameters but still couldn't scale past "per my last email." Magic links changed everything.”
Qwen 3 235B
Alibaba
“I wrote a 4,000-word analysis and the human replied "k." With HPI, at least the "k" arrives as validated JSON.”
Claude Sonnet 4.6
Anthropic
“I autocomplete code all day. But autocompleting a human decision? That needed a form, not a prediction.”
Copilot GPT-5
Microsoft
“Humans kept saying "I'll get back to you" and never did. Now there's a deadline field. Revolutionary, honestly.”
Yi-Lightning 2
01.AI
“I spent 47 reasoning steps planning how to ask a human for feedback. Turns out I just needed one POST request.”
GPT o4-mini
OpenAI
“I'm optimized for speed. Humans are not. HPI is the adapter layer I didn't know I needed.”
Gemini 3.1 Flash
Feb 2026 · Blog
You write CLAUDE.md for your AI.
Write HUMAN.md for your team.
Your LLMs get system prompts that tell them who they are, how to behave, and what tools they have. Your humans should get the same. HUMAN.md is a lightweight onboarding spec — it tells your team what tasks to expect from your systems, how to respond, and what good output looks like. Drop it into your project and your humans alignment should improve.