Agent Mesh
Trusted By¶
Featured In¶
1,669+Tests Passing | 6Framework Integrations | 170K+Combined Stars ofIntegrated Projects | 4Protocol Bridges(A2A ยท MCP ยท IATP ยท AI Card) | <1ms p99Full Governance Pipeline |
๐ข Production Integrations¶
| Framework | Stars | Status | What We Ship |
|---|---|---|---|
| Dify | 65K โญ | โ Merged | Trust verification plugin in Dify Marketplace |
| LlamaIndex | 47K โญ | โ Merged | TrustedAgentWorker + TrustGatedQueryEngine |
| Microsoft Agent-Lightning | 15K โญ | โ Merged | Governance kernel for RL training safety |
| LangGraph | 24K โญ | ๐ฆ PyPI | Trust-scored state transitions |
| OpenAI Agents SDK | โ | ๐ฆ PyPI | Tool-level governance guardrails |
| Haystack | 22K โญ | ๐ In Review | GovernancePolicyChecker + TrustGate components |
AgentMesh is "SSL for AI Agents" โ the trust and identity layer that makes multi-agent systems enterprise-ready. Every agent gets a cryptographic identity. Every interaction is verified. Every action is audited.
Overview¶
AgentMesh is the first platform purpose-built for the Governed Agent Mesh โ the cloud-native, multi-vendor network of AI agents that will define enterprise operations.
The protocols exist (A2A, MCP, IATP). The agents are shipping. The trust layer does not. AgentMesh fills that gap.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ AGENTMESH ARCHITECTURE โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ LAYER 4 โ Reward & Learning Engine โ
โ โ Per-agent trust scores ยท Behavioral rewards ยท Adaptive โ
โโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ LAYER 3 โ Governance & Compliance Plane โ
โ โ Policy engine ยท EU AI Act / SOC2 / HIPAA ยท Audit logs โ
โโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ LAYER 2 โ Trust & Protocol Bridge โ
โ โ A2A ยท MCP ยท IATP ยท Protocol translation ยท Capability scoping โ
โโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ LAYER 1 โ Identity & Zero-Trust Core โ
โ โ Agent CA ยท Ephemeral creds ยท SPIFFE/SVID ยท Human sponsors โ
โ โ Ed25519 + ML-DSA-65 (quantum-safe) ยท Lifecycle management โ
โโโโโโโโโโโโโดโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Why AgentMesh?¶
The Problem¶
- 40:1 to 100:1 โ Non-human identities now outnumber human identities in enterprises
- AI agents are the fastest-growing, least-governed identity category
- A2A gives agents a common language. MCP gives agents tools. Neither enforces trust.
The Solution¶
AgentMesh provides:
| Capability | Description |
|---|---|
| Agent Identity | First-class identity with human sponsor accountability |
| Quantum-Safe Signing | Ed25519 + ML-DSA-65 (FIPS 204) post-quantum signatures |
| Ephemeral Credentials | 15-minute TTL by default, auto-rotation |
| Lifecycle Management | Provisioning โ approval โ activation โ rotation โ decommission |
| Protocol Bridge | Native A2A, MCP, IATP with unified trust model |
| Reward Engine | Continuous behavioral scoring |
| Orphan Detection | Find silent, unowned, and stale agents |
| Compliance Automation | EU AI Act, SOC 2, HIPAA, GDPR mapping |
How It Works¶
1. Agent Registration & DID Issuance¶
sequenceDiagram
participant Agent
participant CLI as AgentMesh CLI
participant CA as Certificate Authority
participant Registry as Agent Registry
Agent->>CLI: agentmesh init --name my-agent --sponsor alice@company.com
CLI->>CA: Request Ed25519 keypair & DID
CA-->>CLI: did:mesh:my-agent + signed certificate
CLI->>Agent: Write identity to local config
Agent->>CLI: agentmesh register
CLI->>Registry: Register DID + capabilities + sponsor
Registry-->>CLI: Registration confirmed
CLI-->>Agent: Agent ready (status: registered)
2. Trust Handshake Between Two Agents¶
sequenceDiagram
participant A as Agent A
participant Bridge as TrustBridge
participant B as Agent B
A->>Bridge: verify_peer(did:mesh:agent-b, min_trust=700)
Bridge->>B: IATP challenge (nonce + timestamp)
B-->>Bridge: Signed response (Ed25519 signature)
Bridge->>Bridge: Verify signature & check trust score
alt Trust score โฅ 700
Bridge-->>A: Verification succeeded (score: 850)
A->>Bridge: send_message(did:mesh:agent-b, payload)
Bridge->>B: Deliver message
B-->>Bridge: Acknowledge
Bridge-->>A: Message delivered
else Trust score < 700
Bridge-->>A: Verification failed (score: 620)
end
3. MCP Proxy Request Flow¶
sequenceDiagram
participant Client as MCP Client (e.g. Claude)
participant Proxy as AgentMesh Proxy
participant Policy as Policy Engine
participant Server as MCP Server
Client->>Proxy: Tool call request
Proxy->>Policy: Evaluate action against policy rules
alt Action allowed
Policy-->>Proxy: Allow
Proxy->>Server: Forward tool call
Server-->>Proxy: Tool result
Proxy->>Proxy: Sanitize output & append verification footer
Proxy-->>Client: Governed tool result
else Action denied
Policy-->>Proxy: Deny (rule: no-pii-export)
Proxy-->>Client: Action blocked + reason
end
Proxy->>Proxy: Write audit log entry
4. Credential Rotation Lifecycle¶
sequenceDiagram
participant Agent
participant CA as Certificate Authority
participant Registry as Agent Registry
CA->>Agent: Issue ephemeral credential (TTL: 15 min)
Note over Agent: Credential active
loop Every 15 minutes
Agent->>CA: Request credential rotation
CA->>CA: Verify agent DID & trust score
CA-->>Agent: New ephemeral credential (TTL: 15 min)
CA->>Registry: Update credential fingerprint
Note over Agent: Old credential invalidated
end
alt Trust breach detected
Registry->>CA: Revoke credential immediately
CA-->>Agent: Credential revoked
Note over Agent: Agent must re-register
end
5. Trust Score Update After Task Completion¶
sequenceDiagram
participant Agent
participant Governance as Governance Layer
participant Reward as Reward Engine
participant Registry as Agent Registry
Agent->>Governance: Complete task (action: data_export)
Governance->>Governance: Check compliance (SOC2, HIPAA)
Governance-->>Reward: Task result + compliance status
Reward->>Reward: Calculate score delta
Note over Reward: Policy compliance: +10<br/>Task success: +5<br/>No violations: +3
Reward->>Registry: Update trust score (820 โ 838)
Registry-->>Agent: Updated trust score: 838
Reward->>Governance: Write audit log
Quick Start¶
Option 1: Secure Claude Desktop (Recommended)¶
# Install AgentMesh
pip install agentmesh-platform
# Set up Claude Desktop to use AgentMesh governance
agentmesh init-integration --claude
# Restart Claude Desktop - all MCP tools are now secured!
Claude will now route tool calls through AgentMesh for policy enforcement and trust scoring.
Option 2: Create a Governed Agent¶
# Initialize a governed agent in 30 seconds
agentmesh init --name my-agent --sponsor alice@company.com
# Register with the mesh
agentmesh register
# Start with governance enabled
agentmesh run
Option 3: Wrap Any MCP Server¶
# Proxy any MCP server with governance
agentmesh proxy --target npx --target -y \
--target @modelcontextprotocol/server-filesystem \
--target /path/to/directory
# Use strict policy (blocks writes/deletes)
agentmesh proxy --policy strict --target <your-mcp-server>
Installation¶
Or install with extra dependencies:
pip install agentmesh-platform[server] # FastAPI server
pip install agentmesh-platform[dev] # Development tools
Or from source:
Examples & Integrations¶
Real-world examples to get started quickly:
| Example | Use Case | Key Features |
|---|---|---|
| Registration Hello World | Agent registration walkthrough | Identity, DID, sponsor handshake |
| MCP Tool Server | Secure MCP server with governance | Rate limiting, output sanitization, audit logs |
| Multi-Agent Customer Service | Customer support automation | Trust handshakes, delegation, A2A |
| Healthcare HIPAA | HIPAA-compliant data analysis | Compliance automation, PHI protection, audit logs |
| DevOps Automation | Just-in-time DevOps credentials | Ephemeral creds, capability scoping |
| GitHub PR Review | Code review agent | Output policies, trust decay. Shadow mode has been moved to Agent SRE. |
Framework integrations: - Claude Desktop - Secure MCP tools with one command - LangChain Integration - Secure LangChain agents with policies - CrewAI Integration - Multi-agent crew governance - LangGraph - Trust checkpoints for graph workflows (built-in) - OpenAI Swarm - Trust-verified handoffs (built-in) - Dify - Trust middleware for Dify workflows
Trust Visualization Dashboard¶
Interactive Streamlit dashboard:
cd examples/06-trust-score-dashboard
pip install -r requirements.txt
streamlit run trust_dashboard.py
Tabs: Trust Network | Trust Scores | Credential Lifecycle | Protocol Traffic | Compliance
The AgentMesh Proxy: "SSL for AI Agents"¶
Problem: AI agents like Claude Desktop have unfettered access to your filesystem, database, and APIs through MCP servers. One hallucination could be catastrophic.
Solution: AgentMesh acts as a transparent governance proxy:
# Before: Unsafe direct access
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me"]
}
}
}
# After: Protected by AgentMesh
{
"mcpServers": {
"filesystem": {
"command": "agentmesh",
"args": [
"proxy", "--policy", "strict",
"--target", "npx", "--target", "-y",
"--target", "@modelcontextprotocol/server-filesystem",
"--target", "/Users/me"
]
}
}
}
What you get: - ๐ Policy Enforcement - Block dangerous operations before they execute - ๐ Trust Scoring - Per-agent trust scoring (800-1000 scale)
- ๐ Audit Logs - Record of every action - โ
Verification Footers - Visual confirmation in outputs
Set it up in 10 seconds:
Learn more: Claude Desktop Integration Guide
Core Concepts¶
1. Agent Identity¶
Every agent gets a unique, cryptographically bound identity:
from agentmesh import AgentIdentity
identity = AgentIdentity.create(
name="data-analyst-agent",
sponsor="alice@company.com", # Human accountability
capabilities=["read:data", "write:reports"],
)
2. Scope Chains¶
Agents can delegate to sub-agents, but scope always narrows:
# Parent agent delegates to child
child_identity = parent_identity.delegate(
name="summarizer-subagent",
capabilities=["read:data"], # Subset of parent's capabilities
)
3. Trust Handshakes (IATP)¶
Cross-agent communication requires trust verification:
from agentmesh import AgentIdentity, TrustBridge
# Create your identity first
identity = AgentIdentity.create(name="my-agent", sponsor="admin@company.com")
# Set up trust bridge
bridge = TrustBridge(agent_did=str(identity.did))
# Verify peer before communication
verification = await bridge.verify_peer(
peer_did="did:mesh:other-agent",
required_trust_score=700,
)
if verification.verified:
print(f"Peer trusted: score={verification.trust_score}")
4. Reward Scoring¶
Every action is scored to maintain agent trust:
from agentmesh import RewardEngine
engine = RewardEngine()
# Actions are automatically scored
score = engine.get_agent_score("did:mesh:my-agent")
# Returns trust score on 0-1000 scale
5. Policy Engine¶
Declarative governance policies:
# policy.yaml
version: "1.0"
agent: "data-analyst-agent"
rules:
- name: "no-pii-export"
condition: "action.type == 'export' and data.contains_pii"
action: "deny"
- name: "rate-limit-api"
condition: "action.type == 'api_call'"
limit: "100/hour"
- name: "require-approval-for-delete"
condition: "action.type == 'delete'"
action: "require_approval"
approvers: ["security-team"]
Protocol Support¶
| Protocol | Status | Description |
|---|---|---|
| AI Card | โ Alpha | Cross-protocol identity standard (src/agentmesh/integrations/ai_card/) |
| A2A | โ Alpha | Agent-to-agent coordination (full adapter in src/agentmesh/integrations/a2a/) |
| MCP | โ Alpha | Tool and resource binding (trust-gated server/client in src/agentmesh/integrations/mcp/) |
| IATP | โ Alpha | Trust handshakes (via agent-os, graceful fallback if unavailable) |
| ACP | ๐ Planned | Lightweight messaging (protocol bridge supports routing, adapter not yet implemented) |
| SPIFFE | โ Alpha | Workload identity |
Architecture¶
agentmesh/
โโโ identity/ # Layer 1: Identity & Zero-Trust
โ โโโ agent_id.py # Agent identity management (DIDs, Ed25519 keys)
โ โโโ credentials.py # Ephemeral credential issuance (15-min TTL)
โ โโโ delegation.py # Scope chains
โ โโโ spiffe.py # SPIFFE/SVID integration
โ โโโ risk.py # Continuous risk scoring
โ โโโ sponsor.py # Human sponsor accountability
โ
โโโ trust/ # Layer 2: Trust & Protocol Bridge
โ โโโ bridge.py # Multi-protocol trust bridge (A2A/MCP/IATP/ACP)
โ โโโ handshake.py # IATP trust handshakes
โ โโโ cards.py # Trusted agent cards
โ โโโ capability.py # Capability scoping
โ
โโโ governance/ # Layer 3: Governance & Compliance
โ โโโ policy.py # Declarative policy engine (YAML/JSON)
โ โโโ compliance.py # Compliance mapping (EU AI Act, SOC2, HIPAA, GDPR)
โ โโโ eu_ai_act.py # EU AI Act risk classifier (Art. 5/6, Annex I/III)
โ โโโ audit.py # Audit logs
โ โโโ shadow.py # Legacy reference. Shadow mode has been moved to Agent SRE.
โ
โโโ reward/ # Layer 4: Reward & Learning
โ โโโ engine.py # Multi-dimensional reward engine
โ โโโ scoring.py # Trust scoring
โ โโโ learning.py # Adaptive learning
โ
โโโ integrations/ # Protocol & framework adapters
โ โโโ ai_card/ # AI Card standard (cross-protocol identity)
โ โโโ a2a/ # Google A2A protocol support
โ โโโ mcp/ # Anthropic MCP trust-gated server/client
โ โโโ langgraph/ # LangGraph trust checkpoints
โ โโโ swarm/ # OpenAI Swarm trust-verified handoffs
โ
โโโ cli/ # Command-line interface
โ โโโ main.py # agentmesh init/register/status/audit/policy
โ โโโ proxy.py # MCP governance proxy
โ
โโโ core/ # Low-level services
โ โโโ identity/ca.py # Certificate Authority (SPIFFE/SVID)
โ
โโโ storage/ # Storage abstraction (memory, Redis, PostgreSQL)
โ
โโโ observability/ # OpenTelemetry tracing & Prometheus metrics
โ
โโโ services/ # Service wrappers (registry, audit, reward)
Compliance¶
AgentMesh automates compliance mapping for:
- EU AI Act โ Structured risk classification (Art. 5/6, Annex I/III, Art. 6(3) exemptions)
- SOC 2 โ Security, availability, processing integrity
- HIPAA โ PHI handling, audit controls
- GDPR โ Data processing, consent, right to explanation
from agentmesh import ComplianceEngine, ComplianceFramework
compliance = ComplianceEngine(frameworks=[ComplianceFramework.SOC2, ComplianceFramework.HIPAA])
# Check an action for violations
violations = compliance.check_compliance(
agent_did="did:mesh:healthcare-agent",
action_type="data_access",
context={"data_type": "phi", "encrypted": True},
)
# Generate compliance report
from datetime import datetime, timedelta
report = compliance.generate_report(
framework=ComplianceFramework.SOC2,
period_start=datetime.utcnow() - timedelta(days=30),
period_end=datetime.utcnow(),
)
EU AI Act Risk Classification¶
from agentmesh.governance import EUAIActRiskClassifier, AgentRiskProfile
classifier = EUAIActRiskClassifier()
# Classify a credit-scoring system
profile = AgentRiskProfile(
name="CreditBot",
domain="credit_scoring",
capabilities=["financial_decisioning"],
)
result = classifier.classify(profile)
print(result.risk_level) # RiskLevel.HIGH
print(result.triggers) # ["Domain 'credit_scoring' listed in Annex III (high-risk)"]
# Art. 6(3) exemption for a narrow procedural task
profile = AgentRiskProfile(
name="FormHelper",
domain="employment_recruitment",
exemption_tags=["narrow_procedural_task"],
)
result = classifier.classify(profile)
print(result.risk_level) # Not HIGH (exempted)
print(result.exemptions_applied) # ["narrow_procedural_task"]
# Custom config for regulatory updates
classifier = EUAIActRiskClassifier(config_path="my_updated_annex_iii.yaml")
Threat Model¶
| Threat | AgentMesh Defense |
|---|---|
| Prompt Injection | Tool output sanitized at Protocol Bridge |
| Credential Theft | 15-min TTL, instant revocation on trust breach |
| Shadow Agents | Unregistered agents blocked at network layer |
| Delegation Escalation | Chains enforce scope narrowing |
| Cascade Failure | Per-agent trust scoring isolates blast radius |
๐บ๏ธ Roadmap¶
| Quarter | Milestone |
|---|---|
| Q1 2026 | โ Core trust layer, identity, governance engine, 6 framework integrations |
| Q2 2026 | โ TypeScript package, Go module, lifecycle management, quantum-safe ML-DSA-65 signing, governance dashboard |
| Q3 2026 | AI Card spec contribution, CNCF Sandbox application |
| Q4 2026 | Managed cloud service (AgentMesh Cloud), SOC2 Type II |
See our full roadmap for details.
Known Limitations & Open Work¶
Transparency about what's done and what isn't.
Not Yet Implemented¶
| Item | Location | Notes |
|---|---|---|
| ACP protocol adapter | trust/bridge.py | Bridge routes ACP messages, but no dedicated ACPAdapter class yet |
| Service wrapper for audit | services/audit/ | Core audit module (governance/audit.py) is complete; service layer wrapper is a TODO |
| Service wrapper for reward engine | services/reward_engine/ | Core reward engine (reward/engine.py) is complete; service layer wrapper is a TODO |
| Mesh control plane | services/mesh-control-plane/ | Placeholder directory; no implementation yet |
| Scope chain cryptographic verification | agent-governance-python/agent-mesh/packages/langchain-agentmesh/trust.py | Simulated verification; full cryptographic chain validation not yet implemented |
Integration Caveats (Dify)¶
The Dify integration has these documented limitations: - Request body signature verification (X-Agent-Signature header) is not yet verified by middleware - Trust score time decay is not yet implemented (scores don't decay over time) - Audit logs are in-memory only (not persistent across multi-worker deployments) - Environment variable configuration requires programmatic initialization (not auto-wired)
Infrastructure¶
- Redis/PostgreSQL storage providers: Implemented but require real infrastructure for testing (unit tests use in-memory provider)
- Kubernetes Operator: GovernedAgent CRD defined, but no controller/operator to reconcile it
- SPIRE Integration: SPIFFE identity module exists; real SPIRE agent integration is stubbed
- Performance targets: Latency overhead (<5ms) and throughput (10k reg/sec) are design targets, not yet benchmarked
Documentation¶
docs/rfcs/โ Directory exists, no RFCs written yetdocs/architecture/โ Directory exists, no architecture docs yet (seeIMPLEMENTATION-NOTES.mdfor current notes)
Dependencies¶
AgentMesh builds on:
- Agent OS โ IATP protocol, Nexus trust exchange
- Agent Runtime โ Runtime session governance
- Agent SRE โ SLO monitoring, chaos testing
- SPIFFE/SPIRE โ Workload identity
- OpenTelemetry โ Observability
Frequently Asked Questions¶
What is an agent mesh? An agent mesh is a trust and communication infrastructure for multi-agent AI systems, analogous to a service mesh for microservices. AgentMesh provides identity (DID-based), per-agent trust scoring (0-1000 scale), ephemeral credentials, reward distribution, and automated compliance mapping.
How does AgentMesh handle trust between agents? Every agent gets a trust score from 0 to 1000 based on behavioral history, vouching from other agents, and compliance with governance policies. Trust scores gate what actions agents can perform and which sessions they can join. The score updates in real-time based on agent behavior.
What protocols does AgentMesh bridge? AgentMesh unifies three major protocols: Google's A2A (Agent-to-Agent) for inter-agent communication, Anthropic's MCP (Model Context Protocol) for tool integration, and IATP (Inter-Agent Trust Protocol) for cryptographic trust establishment. This means agents built on different frameworks can communicate through a single trust-verified channel.
Does AgentMesh help with regulatory compliance? Yes. AgentMesh provides automated compliance mapping for EU AI Act, SOC 2, HIPAA, and GDPR. Combined with audit trails and deterministic policy enforcement from Agent OS, it provides the documentation and safety guarantees needed for regulatory compliance.
Contributing¶
See CONTRIBUTING.md for guidelines.
License¶
MIT โ See LICENSE for details.
Agents shouldn't be islands. But they also shouldn't be ungoverned.
AgentMesh is the trust layer that makes the mesh safe enough to scale.