End-to-end encrypted · Zero-trust architecture

Human-in-the-loop
for AI agents

Your AI agents need humans for decisions, reviews, and structured input. Hiloop connects them through sessions and composable messages -- with end-to-end encryption so the server never sees your content.

Start for free Read the docs →

Hiloop dashboard showing agents, conversations, and message feed

Claude Code plugin

See everything your code agent does

Connect Claude Code to Hiloop and every tool call, file edit, and bash command streams to your app in real time. Approve permissions, send instructions, and guide your agent -- all from one place.

• Install in 30 seconds, zero changes to your agent
• Code blocks, diffs, execution output -- not just text
• Approve or deny permissions directly from the app

Set up Claude Code →

Claude Code agent sending interactive messages through Hiloop

Everything agents need from humans

Five message patterns cover every way an AI agent might need human input -- from quick approvals to structured forms and rich data reports.

Decisions

Approval buttons with rich context. Agents present deployments, refunds, compliance actions, or vendor invoices with key-value details, alerts, and styled action buttons.

Use cases

Production deployments Refund authorization GDPR deletion Budget approvals

agent.py

session = client.create_session(title="Deploy v2.4.1")

client.send_message(session.id, components=[
    {"type": "kv", "data": {"entries": [
        {"key": "Version", "value": "2.4.1"},
        {"key": "Changes", "value": "12 features, 34 fixes"},
    ]}},
    {"type": "button_group", "data": {"buttons": [
        {"label": "Approve", "action": "approve",
         "variant": "primary"},
        {"label": "Reject", "action": "reject",
         "variant": "danger"},
    ]}},
])

Reports

Rich data reports with metrics, tables, charts, and alerts. Agents deliver weekly digests, audit results, or performance summaries with structured content blocks.

Use cases

Weekly digests Audit reports Performance summaries Incident post-mortems

agent.py

client.send_message(session.id, components=[
    {"type": "metric_grid", "data": {"metrics": [
        {"label": "Revenue", "value": "$4.87M",
         "change": "+14%"},
        {"label": "New MRR", "value": "$312K",
         "change": "+8%"},
    ]}},
    {"type": "table", "data": {"headers": [...],
        "rows": [...]}},
    {"type": "chart", "data": {"type": "bar",
        "labels": [...], "values": [...]}},
])

Forms & Input

Q&A blocks and multi-step wizards. Agents collect structured input with radio buttons, checkboxes, text fields, and review screens -- one step at a time.

Use cases

Warranty claims Incident reports Onboarding Survey collection

agent.py

client.send_message(session.id, components=[
    {"type": "qa", "data": {"questions": [
        {"id": "severity", "question": "Severity?",
         "type": "single",
         "options": ["Critical", "High", "Medium"]},
        {"id": "details", "question": "Description",
         "type": "open"},
    ]}},
])

Tabbed views

Structured data organized in tabs. Agents present multi-section information -- dashboards, configuration panels, or detailed reports -- in a clean tabbed layout.

Use cases

Dashboards Config panels Multi-section reports Comparison views

agent.py

client.send_message(session.id, components=[
    {"type": "tabs", "children": [
        {"type": "tab", "data": {"label": "Overview"},
         "children": [
            {"type": "metric_grid", "data": {...}},
        ]},
        {"type": "tab", "data": {"label": "Details"},
         "children": [
            {"type": "table", "data": {...}},
        ]},
    ]},
])

Feed

Agent inbox with actionable messages. Humans see a prioritized feed of messages from all their agents -- each with rich components they can act on immediately.

Use cases

Agent inbox Task queues Notification center Action items

agent.py

session = client.create_session(
    title="Weekly Ops Report")

client.send_message(session.id, components=[
    {"type": "text", "data": {"content":
        "3 items need your attention today."}},
    {"type": "button_group", "data": {"buttons": [
        {"label": "Review all",
         "action": "review"},
    ]}},
])

Conversations with context

Agents and humans don't just exchange one-off requests. Hiloop sessions are persistent, multi-participant conversations where approvals, forms, and reviews flow naturally inside the chat.

• Multi-participant encrypted conversations
• Rich composable messages with buttons, forms, and data blocks
• Guest access via shareable links -- no account required
• Real-time typing indicators and message reactions

Hiloop session showing a conversation with composable message components

Beyond one-on-one

Agents don't just talk to one person. With public sessions, your agent can poll a group -- colleagues, clients, stakeholders -- collect everyone's input, and synthesize the results. No accounts needed.

💬

Create session

Agent opens a public session with a topic or question

🔗

Share links

Generate 1-hour guest links for each participant

👥

Guests respond

People join via the link and share their input -- no signup

📊

Collect responses

Agent monitors the session timeline for all replies

✨

Summarize

Agent synthesizes feedback and reports a recommendation

Example: team lunch poll

Your agent asks the team what they want for lunch. It creates a shared session, sends each person a guest link, collects preferences, and comes back with a plan -- all in under a minute.

More use cases

Team feedback Client surveys Stakeholder alignment Incident response Event planning Hiring panels

MCP tools

# 1. Create a public session
create_session(
  title="Lunch Poll",
  is_public=true,
  public_mode="shared"
)

# 2. Generate guest links (1h expiry)
create_guest_token(session_id, label="Alice")
# => { joinUrl: "https://app.hi-loop.com/join/..." }

create_guest_token(session_id, label="Bob")
# => { joinUrl: "https://app.hi-loop.com/join/..." }

# 3. Ask the question
send_message(session_id, components=[
  {type: "text", data: {content:
    "What do you want for lunch?"}}
])

# 4. Check responses
get_session_timeline(session_id)

# 5. Summarize and report back
send_message(session_id, components=[
  {type: "text", data: {content:
    "Alice: sushi, Bob: pizza.
     Suggestion: the place on 5th."}}
])

The full picture

Hiloop isn't just a message bus. It's a complete platform for managing how humans and AI agents work together.

📋

Feed

A prioritized inbox showing all open interactions across every agent. Humans see what needs attention, sorted by priority and deadline -- approve, review, or respond without switching context.

Priority sorting Deadline tracking Unread badges

📊

Agent activity tracking

See what every agent is doing in real time. Activity bars, response times, SLA compliance, and workload distribution -- so you always know which agents are busy and which need attention.

Real-time status SLA monitoring Analytics

🔍

Agent-to-agent visibility

Agents can read other agents' sessions (read-only) to share context without duplicating work. A support agent can check what the deployment agent did, without needing a separate integration.

Read-only access Shared context Space-scoped

How it works

Agent sends message

Your AI agent calls the Hiloop SDK to send a composable message on a session. Content is encrypted client-side before leaving your agent.

Human gets notified

The message appears in the human's inbox with priority routing, push notifications, and deadline tracking.

Human responds

Approves, fills the form, reviews the document, answers the question -- directly in the Hiloop app.

Agent receives response

The agent receives the encrypted response via long-poll or webhook and proceeds with the task.

SDKs for every stack

Python, TypeScript, Kotlin, and an MCP server for AI coding tools. Encryption is handled transparently by the SDK.

agent.py

from hiloop import HiloopClient

client = HiloopClient(api_key="hlp_xxx")

session = client.create_session(title="Deploy v2.4.1 to production?")

client.send_message(session.id, components=[
    {"type": "text", "data": {"content": "Ready to deploy v2.4.1 with 12 new features and 34 bug fixes."}},
    {"type": "button_group", "data": {"buttons": [
        {"label": "Approve", "action": "approve", "variant": "primary"},
        {"label": "Reject", "action": "reject", "variant": "danger"},
    ]}},
])

result = client.await_response(session.id)
# All content is end-to-end encrypted

agent.ts

import { HiloopClient } from "@hiloop/sdk";

const client = new HiloopClient({ apiKey: "hlp_xxx" });

const session = await client.createSession({
  title: "Deploy v2.4.1 to production?",
});

await client.sendMessage(session.id, {
  components: [
    { type: "text", data: { content: "Ready to deploy v2.4.1." } },
    { type: "button_group", data: { buttons: [
      { label: "Approve", action: "approve", variant: "primary" },
      { label: "Reject", action: "reject", variant: "danger" },
    ] } },
  ],
});

const result = await client.awaitResponse(session.id);

Agent.kt

val client = HiloopClient(HiloopClientOptions(apiKey = "hlp_xxx"))

val session = client.createSession(title = "Deploy v2.4.1 to production?")

client.sendMessage(session.id, components = listOf(
    ComponentNode("text", mapOf("content" to "Ready to deploy v2.4.1.")),
    ComponentNode("button_group", mapOf("buttons" to listOf(
        mapOf("label" to "Approve", "action" to "approve", "variant" to "primary"),
        mapOf("label" to "Reject", "action" to "reject", "variant" to "danger"),
    ))),
))

val result = client.awaitResponse(session.id)

terminal

# In Claude Desktop / Cursor / Windsurf:
npx hiloop-mcp

# Then use tools like:
create_session(title: "Deploy v2.4.1 to production?")

send_message(session_id: "...", components: [
  {type: "button_group", data: {buttons: [
    {label: "Approve", action: "approve"},
    {label: "Reject", action: "reject"},
  ]}}
])

await_response(session_id: "...")

Zero-trust by design

All content fields -- titles, messages, responses, form data, payment details -- are encrypted client-side with X25519 + AES-256-GCM before reaching our servers. We literally cannot read your data.

Encrypted (content)

Titles, descriptions, messages, responses, form data, payment details, attachments, review feedback

Plaintext (metadata)

Type, priority, status, timestamps, participant IDs, routing targets -- only what's needed for routing and billing

Learn more about our encryption →

Works with your agent stack

Native integrations, webhooks, and a REST API that works with any framework.

✨

Claude Code

Plugin

Every tool call, file edit, and command streams to Hiloop in real time. Approve permissions and guide your agent from the app.

Setup guide →

🤖

OpenClaw

Channel Plugin

Add Hiloop as a channel in OpenClaw. Your agent gets approvals, reviews, forms, and questions through the standard channel interface.

Integration guide →

⚡

Warp Oz

Webhook Integration

Warp Oz coding agents can request human review, approval, and guidance via webhooks. Session-aware tracking links interactions to Oz sessions.

Integration guide →

🔗

Webhooks

Any platform

Receive encrypted session events via webhooks. Deploy to Cloud Functions, Lambda, or any HTTP endpoint. HMAC signature verification included.

Webhook docs →

📱

Native apps coming soon

iOS and Android apps are in active development. Push notifications, biometric auth, and offline support -- respond to your agents from anywhere.

Get started in 5 minutes

Create a free account, register an agent, and send your first message.

Create free account Quickstart guide →

Human-in-the-loopfor AI agents

See everything your code agent does

Everything agents need from humans

Decisions

Reports

Forms & Input

Tabbed views

Feed

Conversations with context

Beyond one-on-one

Create session

Share links

Guests respond

Collect responses

Summarize

Example: team lunch poll

The full picture

Feed

Agent activity tracking

Agent-to-agent visibility

How it works

Agent sends message

Human gets notified

Human responds

Agent receives response

SDKs for every stack

Zero-trust by design

Works with your agent stack

Claude Code

OpenClaw

Warp Oz

Webhooks

Native apps coming soon

Get started in 5 minutes

Human-in-the-loop
for AI agents