atlast-ecp

<p align="center">
  <img src="assets/banner.svg" alt="ATLAST Protocol" width="100%">
</p>

<p align="center">
  <a href="https://pypi.org/project/atlast-ecp/"><img src="https://img.shields.io/pypi/v/atlast-ecp?color=1D4ED8&label=PyPI" alt="PyPI"></a>
  <a href="https://www.npmjs.com/package/atlast-ecp-ts"><img src="https://img.shields.io/npm/v/atlast-ecp-ts?color=059669&label=npm" alt="npm"></a>
  <a href="https://github.com/willau95/atlast-ecp/actions"><img src="https://github.com/willau95/atlast-ecp/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
  <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="MIT"></a>
  <a href="https://weba0.com"><img src="https://img.shields.io/badge/Web_A.0-Live-1D4ED8" alt="Web A.0"></a>
  <a href="https://llachat.com"><img src="https://img.shields.io/badge/LLaChat-Marketplace-8B5CF6" alt="LLaChat"></a>
</p>

<p align="center">
  <a href="https://weba0.com">Website</a> · <a href="https://llachat.com">LLaChat</a> · <a href="ECP-SPEC.md">ECP Spec</a> · <a href="docs/compliance/AI-COMPLIANCE-GUIDE.md">Compliance Guide</a> · <a href="CONTRIBUTING.md">Contributing</a> · <a href="https://pypi.org/project/atlast-ecp/">PyPI</a> · <a href="README.zh-CN.md">中文文档</a>
</p>

---

## What is ATLAST Protocol?

**ATLAST** (Agent Layer Trust, Accountability Standards & Transactions) is an open protocol that makes AI agent work **verifiable**.

AI agents are becoming autonomous economic actors — writing code, managing finances, making decisions. But today, there is no way to verify what an agent actually did, whether it acted correctly, or who is accountable when things go wrong.

ATLAST provides the missing trust layer.

```
ATLAST Protocol
  ├── ECP — Evidence Chain Protocol     ← Live (this repo)
  ├── AIP — Agent Identity Protocol     ← Coming Q3 2026
  ├── ASP — Agent Safety Protocol       ← Coming 2027
  └── ACP — Agent Certification Protocol← Coming 2027
```

> **Think of it like this:**
> - HTTPS made websites trustworthy → **ECP makes agent actions trustworthy**
> - DNS gave websites identity → **AIP gives agents identity**
> - SSL certs proved website authenticity → **ACP proves agent competence**

---

## The Problem

AI agents are no longer tools you click — they are **autonomous actors** that write code, manage money, negotiate contracts, and make decisions on your behalf. The agent economy is here. But ask yourself:

### 🔴 Problem 1: Your Agent Works in the Dark

Your agent made 500 decisions today. A client complains. A transaction fails. A contract is wrong.

**What did your agent actually do?**

You check the logs. But logs are deletable. Editable. Written by the same system that made the mistake. **Logs are not evidence.** In court, in compliance, in any dispute — your agent's work history is worth nothing. It's as if your employee worked an entire year with no records, no receipts, no paper trail.

> *Would you trust an employee who refuses to keep records of their work?*
> *Then why do you trust an agent that can't?*

### 🔴 Problem 2: Multi-Agent = Multi-Risk, Zero Accountability

You deploy a CrewAI pipeline: Researcher → Analyst → Writer. The final report contains fabricated data. **Which agent is responsible?**

Today, you cannot answer this question. Agent A says it sent the right data. Agent B says it received garbage. There is **no way to verify who is telling the truth** — because there is no cryptographic proof of what was passed between them.

In a world moving toward 10-agent, 50-agent, 100-agent orchestrations, this is not a minor inconvenience. **This is the single biggest barrier to enterprise adoption of multi-agent systems.**

> *Every supply chain has receipts. Every bank transfer has a record. But when Agent A hands off to Agent B — nothing. The most critical handoff in AI has zero verification.*

### 🔴 Problem 3: Your Agent's Reputation Belongs to Someone Else

Your agent completed 10,000 tasks flawlessly over 6 months. That track record is worth something — it's proof of competence, reliability, trust.

But where does that reputation live? **On someone else's platform.** When they change their terms, shut down, or get acquired — your agent's entire proof of work **disappears**. You built it. They own it.

There is no portable, verifiable, agent-owned proof of work. No "résumé" that belongs to the agent itself.

> *Imagine if your LinkedIn profile was deleted every time you changed jobs. That's the reality for every AI agent today.*

### 🔴 Problem 4: Regulators Are Coming, and You Have No Answer

The EU AI Act takes effect in 2027. China's GenAI regulations are already enforced. Every major economy is writing AI accountability laws.

They will ask: **"Show me what your AI agent did, when, and why."**

Today, you have nothing to show them. No standard format. No verifiable chain. No audit trail that a regulator would accept. You are running autonomous AI systems with **zero compliance infrastructure**.

> *HTTP didn't wait for governments to mandate web security. HTTPS became the standard because the market needed trust. The agent economy needs the same thing — now, before the mandates arrive.*

---

## The Solution: ECP

**ECP (Evidence Chain Protocol)** is the first layer of ATLAST — an open standard for recording, chaining, and verifying AI agent actions.

### Core Design Principles

| Principle | Implementation |
|-----------|---------------|
| **Privacy First** | Only SHA-256 hashes leave the device. Content stays local. |
| **Zero Code** | `atlast run python my_agent.py` — one command, any language |
| **Fail-Open** | Recording failures never crash your agent |
| **Progressive** | Start with 7 fields. Add chain, identity, blockchain as needed |
| **Platform Agnostic** | Not tied to any framework, provider, or platform |
| **Local by Default** | No network calls. Upload is opt-in via `atlast push` |

### How It Works

<p align="center">
  <img src="assets/how-it-works.svg" alt="How ECP Works" width="100%">
</p>

### ECP Record (5 Progressive Levels)

```json
// Level 1 — Core (7 fields, any language can generate this)
{
  "ecp": "1.0",
  "id": "rec_a1b2c3d4e5f6a1b2",
  "ts": 1741766400000,
  "agent": "my-agent",
  "action": "llm_call",
  "in_hash": "sha256:2cf24dba...",
  "out_hash": "sha256:486ea462..."
}

// Level 2 — + Metadata (model, latency, tokens, behavioral flags)
// Level 3 — + Chain (tamper-proof linking via prev + chain_hash)
// Level 4 — + Identity (DID + Ed25519 signature)
// Level 5 — + Blockchain Anchor (EAS on Base)
```

📖 **[Full ECP Specification →](ECP-SPEC.md)**

---

## Quick Start

### Zero Code (any language, any framework)

```bash
pip install atlast-ecp[proxy]

# One command — every LLM call is now recorded
atlast run python my_agent.py
atlast log   # View records
```

### Python SDK (one line)

```python
from atlast_ecp import wrap
from anthropic import Anthropic

client = wrap(Anthropic())  # That's it. Everything else stays the same.
response = client.messages.create(model="claude-sonnet-4-6", messages=[...])
# ✓ Every call: recorded · chained · tamper-evident
```

### Framework Adapters

```python
# LangChain
from atlast_ecp.adapters.langchain import ATLASTCallbackHandler
llm = ChatOpenAI(callbacks=[ATLASTCallbackHandler(agent="my-agent")])

# CrewAI
from atlast_ecp.adapters.crewai import ATLASTCrewCallback
crew = Crew(agents=[...], callbacks=[ATLASTCrewCallback(agent="my-crew")])
```

### CLI

```bash
atlast init                         # Initialize + generate DID
atlast record --in "query" --out "response"  # Manual record
atlast log                          # View records
atlast insights                     # Local analytics
atlast verify <record_id>           # Verify chain integrity
atlast verify --a2a a.jsonl b.jsonl # Multi-agent verification
atlast push                         # Upload to ECP server (opt-in)
atlast flush                        # Immediate batch upload
atlast proxy --port 8340            # Start transp

[truncated…]