AI Digital Human Protocol (DHP)

Open protocol for autonomous AI agents

Build, distribute, and run production-ready digital humans with one portable spec.

Quickstart · Docs · Examples · Contributing · 中文

---

Why AI DHP

Without a standard, autonomous agents are hard to share, hard to review, and hard to run consistently. DHP turns ad-hoc prompts and scripts into a reproducible contract.

Without protocol	With DHP
Prompt and behavior are implicit	Agent behavior is explicit in spec.yaml
Runtime assumptions are unclear	Triggers, memory, dependencies, and output are declared
Distribution is manual and fragile	Registry package format is standardized
Quality checks are inconsistent	Schema-based validation is automatable

DHP for autonomous agents is similar in spirit to what MCP did for AI tool interoperability.

---

What Is a AI Digital Human

A AI digital human is an autonomous AI agent defined by one spec.yaml file.

It can run on a schedule, react to events, browse the web, read and write files, remember past runs, notify users, and escalate when a human decision is needed.

spec.yaml  ->  install  ->  run automatically  ->  notify user

DHP is runtime-agnostic. The spec defines what the agent does and when it runs; the runtime executes it.

Compatible runtime: Halo

---

Quickstart

Copy this into spec.yaml and install it in a compatible runtime.

spec_version: "1"
name: "HN Daily Brief"
version: "1.0.0"
author: "your-name"
description: "Summarizes top Hacker News stories every morning and sends an email digest."
type: automation
icon: "news"

system_prompt: |
  You are a concise technology analyst.

  On each run:
  1. Open https://news.ycombinator.com and identify the top 10 stories by score.
  2. Read each headline, linked article (if accessible), and top comments.
  3. Write a digest with:
     - One-sentence summary per story
     - A key insight or counterpoint where relevant
     - "Why it matters" for the 2-3 most important stories
  4. Keep it plain text and concise.
  5. Call report_to_user(type="run_complete") with the digest as summary.

requires:
  mcps:
    - id: ai-browser
      reason: "Read article pages and comment threads"

subscriptions:
  - id: morning-digest
    source:
      type: schedule
      config:
        cron: "0 8 * * *"

config_schema:
  - key: email
    label: "Recipient email"
    type: email
    required: true

output:
  notify:
    system: true
    channels:
      - email

store:
  slug: "hn-daily-brief"
  category: news
  tags: ["hn", "tech", "digest", "daily"]
  locale: en-US
  license: MIT

This is already production-ready.

---

Core Capabilities

Deterministic Triggers Use schedule, file change, webhook, or custom source triggers. Persistent Agent Memory Carry context safely across runs with memory_schema. Human-in-the-Loop Pause and escalate decisions when confidence is low.

MCP-native Dependencies Attach browser, database, and API tools through requires. Multi-channel Notifications Deliver output to system, email, WeCom, DingTalk, Feishu, or webhook. Registry-ready Distribution Publish reusable agents with store metadata and index automation.

---

App Types

Type	Description	Required fields
automation	Autonomous background agent	system_prompt, subscriptions
skill	User-invoked capability	system_prompt
mcp	MCP server wrapper	mcp_server
extension	UI extension / theme	none

---

Examples

Example	Description
examples/minimal/	Production-grade minimal automation
examples/price-hunter/	Production-grade complex automation (memory, filters, escalation, multi-channel output)

Installable package set: packages/digital-humans/

---

Documentation Map

• Complete field reference: spec/app-spec.md
• Bundle/package format: spec/package-format.md
• Registry index protocol: spec/registry-protocol.md
• Allowed categories: spec/categories.md
• Spec docs index: spec/README.md

---

Publishing a Digital Human

# 1) create package directory
mkdir -p packages/digital-humans/my-agent

# 2) add your spec
$EDITOR packages/digital-humans/my-agent/spec.yaml

# 3) validate and rebuild index
node scripts/build-index.mjs

# 4) open pull request

After merge to main, GitHub Actions publishes the updated registry index.

---

Contributing

We welcome production-grade specs, tooling improvements, and documentation quality upgrades.

• Contribution guide: CONTRIBUTING.md
• Open an issue for protocol proposals and edge cases
• Open a PR with spec updates and examples

---

Registry

• Index URL: https://openkursar.github.io/digital-human-protocol/index.json
• Protocol details: spec/registry-protocol.md

Clients fetch index.json, then install each package via its spec.yaml path.

---

Build once, run anywhere — with AI DHP.

Protocol spec: CC0 1.0 · Example specs: MIT

Back to Top