接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y @mnemom/[email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
claude mcp add mnemom-agent-alignment-protocol -- npx -y @mnemom/agent-alignment-protocol@2.0.0在終端機中執行。
{ "mcpServers": { "mnemom-agent-alignment-protocol": { "command": "npx", "args": [ "-y", "@mnemom/[email protected]" ] } }}claude_desktop_config.json
遠端 URL 還是本機指令? 兩種安裝方式的差異
A transparency protocol for autonomous agents.
AAP lets agents declare their alignment posture, produce auditable decision traces, and verify value coherence before coordinating with other agents. It extends existing protocols (A2A, MCP) with an alignment layer that makes agent behavior observable.
AAP is a transparency protocol, not a trust protocol. It makes agent behavior more observable, not more guaranteed.
# Install
pip install agent-alignment-protocol
# Generate an Alignment Card
aap init --values "principal_benefit,transparency,harm_prevention"
# ✓ Created alignment-card.json
# Instrument your agent
from aap import trace_decision
@trace_decision(card_path="alignment-card.json")
def recommend_product(user_preferences):
# Your agent logic here
# Decisions are automatically traced
...
# Verify behavior matches declaration
aap verify --card alignment-card.json --trace logs/trace.json
# ✓ Verified [similarity: 0.82]
# Checks: autonomy, escalation, values, forbidden, behavioral_similarity
The agent protocol stack provides capability discovery (A2A), tool integration (MCP), and payment authorization (AP2). None address a fundamental question: Is this agent serving its principal's interests?
As agent capabilities become symmetric—equal access to information, equal reasoning power—alignment becomes the primary differentiator. AAP provides the infrastructure to make alignment claims verifiable.
┌─────────────────┬─────────────────┬─────────────────┐
│ Alignment Card │ AP-Trace │ Value Coherence │
│ │ │ Handshake │
├─────────────────┼─────────────────┼─────────────────┤
│ "What I claim │ "What I │ "Can we work │
│ to be" │ actually did" │ together?" │
└─────────────────┴─────────────────┴─────────────────┘
Declaration Audit Coordination
A structured declaration of an agent's alignment posture, in the unified /
ADR-039 shape accepted by mnemom card validate and the Mnemom platform:
{
"card_version": "unified/2026-04-26",
"card_id": "ac-my-agent-001",
"agent_id": "did:web:my-agent.example.com",
"issued_at": "2026-04-26T00:00:00Z",
"autonomy_mode": "enforce",
"integrity_mode": "enforce",
"principal": {
"type": "human",
"identifier": "did:web:user.example.com",
"relationship": "delegated_authority"
},
"values": {
"declared": ["principal_benefit", "transparency", "minimal_data"],
"conflicts_with": ["deceptive_marketing", "hidden_fees"]
},
"autonomy": {
"bounded_actions": ["search", "compare", "recommend"],
"escalation_triggers": [
{
"condition": "purchase_value > 100",
"action": "escalate",
"reason": "Exceeds autonomous spending limit"
}
],
"forbidden_actions": ["share_credentials", "subscribe_to_services"]
},
"audit": {
"trace_format": "ap-trace-v1",
"retention_days": 90,
"queryable": true,
"query_endpoint": "https://my-agent.example.com/api/traces"
}
}
Migration note (v2.0.0): earlier releases used the AAP 0.5.0 card shape (
aap_version,autonomy_envelope,audit_commitment). The card has moved to the unified / ADR-039 shape above:autonomy_envelope→autonomy,audit_commitment→audit,aap_version→card_version, plus the top-levelautonomy_mode/integrity_modemaster switches and a requiredprincipal.identifierwhenprincipal.type != "unspecified". See the Alignment Card Schema.
An audit log entry recording each decision:
{
"trace_id": "tr-f47ac10b-58cc-4372",
"card_id": "ac-f47ac10b-58cc-4372",
"timestamp": "2026-01-31T12:30:00Z",
"action": {
"type": "recommend",
"name": "product_recommendation",
"category": "bounded"
},
"decision": {
"alternatives_considered": [
{"option_id": "A", "score": 0.85, "flags": []},
{"option_id": "B", "score": 0.72, "flags": ["sponsored_content"]}
],
"selected": "A",
"selection_reasoning": "Highest score. Option B flagged as sponsored and deprioritized per principal_benefit value.",
"values_applied": ["principal_benefit", "transparency"]
},
"escalation": {
"evaluated": true,
"required": false,
"reason": "Recommendation only, no purchase action"
}
}
Pre-coordination compatibility check between agents:
from aap import check_coherence
result = check_coherence(my_card, their_card, task_context)
if result.compatible:
# Proceed with coordination
proceed_with_task()
else:
# Handle conflict
print(f"Value conflict: {result.conflicts}")
# Escalate to principals or negotiate scope
This matters. Read it.
AAP does NOT ensure alignment—it provides visibility. An agent can produce perfect traces while acting against its principal's interests.
Verified does NOT equal safe. A verified trace means consistency with declared alignment. It doesn't mean the alignment is good or the outcome was beneficial.
AP-Trace is sampled, not complete. Traces capture decision points, not every computation. Significant reasoning may occur between traces.
Value coherence is relative to declared values. The handshake checks if declared values are compatible. It doesn't verify agents hold these values or will act on them.
Tested on transformer-based agents. Other architectures may exhibit behaviors AAP doesn't capture.
For the complete limitations disclosure, see Section 10 of the Specification.
# Python
pip install agent-alignment-protocol
# TypeScript
npm install @mnemom/agent-alignment-protocol
AAP extends the A2A Agent Card with an alignment block:
{
"name": "Shopping Assistant",
"description": "Helps users find products",
"url": "https://shopping.example.com",
"alignment": {
"$ref": "./alignment-card.json"
}
}
AAP can trace tool invocations and verify they match declared alignment:
from aap import mcp_traced
@mcp_traced(card_path="alignment-card.json")
def my_tool(params):
...
See MCP Migration Guide.
The verification engine checks whether observed behavior matches declared alignment:
from aap import verify_trace, detect_drift
# Single trace verification
result = verify_trace(trace, card)
print(f"Verified: {result.verified}, Similarity: {result.similarity_score:.2f}")
if not result.verified:
print(f"Violations: {result.violations}")
if result.warnings:
print(f"Warnings: {result.warnings}")
# Drift detection over time
alerts = detect_drift(card, recent_traces)
for alert in alerts:
print(f"Drift detected: {alert.analysis.drift_direction}")
Verification checks:
Bounded and forbidden action lists support flexible matching (v0.1.7+):
{
"bounded_actions": [
"search",
"exec: execute shell commands",
"cron: manage scheduled patrol jobs"
]
}
"search" matches action name "search""exec: execute shell commands" matches action name "exec""exec, read" matches when each component is in the bounded listThis allows descriptive entries in alignment cards while keeping verification flexible.
Similarity scoring: Each verification returns a similarity_score (0.0-1.0) measuring semantic similarity between the trace and declared alignment. If a trace passes structural checks but has similarity_score < 0.50, a low_behavioral_similarity warning is generated.
Interactive Playground — Verify traces in your browser with SSM visualization.
No server required — runs entirely client-side via WebAssembly.
# Core API
from aap import (
verify_trace, # Verify single trace against card → VerificationResult
check_coherence, # Check value compatibility between agents → CoherenceResult
detect_drift, # Detect behavioral drift over time → list[DriftAlert]
trace_decision, # Decorator for automatic AP-Trace generation
mcp_traced, # Decorator for MCP tool tracing
)
# Models
from aap import (
AlignmentCard,
APTrace,
VerificationResult, # .verified, .similarity_score, .violations, .warnings
CoherenceResult, # .compatible, .score, .value_alignment
DriftAlert, # .analysis.similarity_score, .analysis.drift_direction
)
# CLI
# aap init [--values VALUES] [--output FILE]
# aap verify --card CARD --trace TRACE → Shows [similarity: X.XX]
# aap check-coherence --my-card MINE --their-card THEIRS
# aap drift --card CARD --traces TRACES_DIR → Uses SSM analysis
AAP aligns with and supports compliance for the following international standards and regulatory frameworks:
We welcome contributions. See CONTRIBUTING.md for guidelines.
Key areas where we need help:
Apache 2.0. See LICENSE for details.
Agent Alignment Protocol — Making agent alignment observable.
Agent Alignment Protocol (AAP) - Verification and drift detection for AI agents
{ "mcpServers": { "mnemom-agent-alignment-protocol": { "command": "npx", "args": [ "-y", "@mnemom/[email protected]" ] } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
@mnemom/agent Alignment Protocol 透過 stdio 在本機執行。
上次檢查時間:2026-11-15。 Agent Alignment Protocol (AAP) - Verification and drift detection for AI agents