Skip to content

README.md

Latest commit

 

History

History
264 lines (202 loc) · 7.73 KB

README.md

File metadata and controls

264 lines (202 loc) · 7.73 KB

AgentMesh Rust SDK

Rust SDK for the Agent Governance Toolkit — policy evaluation, trust scoring, hash-chain audit logging, and Ed25519 agent identity.

Public Preview — APIs may change before 1.0.

Install

Full SDK

cargo add agentmesh
[dependencies]
agentmesh = "3.0.2"

Standalone MCP Package

If you only need the MCP governance/security surface, install the standalone crate:

cargo add agentmesh-mcp
[dependencies]
agentmesh-mcp = "3.0.2"

Quick Start

use agentmesh::{AgentMeshClient, ClientOptions, PolicyDecision};

fn main() {
    // Create a client with a policy
    let opts = ClientOptions {
        policy_yaml: Some(r#"
version: "1.0"
agent: my-agent
policies:
  - name: capability-gate
    type: capability
    allowed_actions: ["data.read", "data.write"]
    denied_actions: ["shell:*"]
  - name: deploy-approval
    type: approval
    actions: ["deploy.*"]
    min_approvals: 2
"#.to_string()),
        ..Default::default()
    };

    let client = AgentMeshClient::with_options("my-agent", opts)
        .expect("failed to create client");

    // Run an action through the governance pipeline
    let result = client.execute_with_governance("data.read", None);
    println!("Decision: {:?}, Allowed: {}", result.decision, result.allowed);

    // Shell commands are denied
    let result = client.execute_with_governance("shell:rm", None);
    assert!(!result.allowed);

    // Audit chain is verifiable
    assert!(client.audit.verify());
}

MCP-Only Quick Start

use agentmesh_mcp::{
    CredentialRedactor, InMemoryNonceStore, McpMessageSigner, SystemClock,
    SystemNonceGenerator,
};
use std::sync::Arc;
use std::time::Duration;

let signer = McpMessageSigner::new(
    b"top-secret-signing-key".to_vec(),
    Arc::new(SystemClock),
    Arc::new(SystemNonceGenerator),
    Arc::new(InMemoryNonceStore::default()),
    Duration::from_secs(300),
    Duration::from_secs(600),
)?;

let message = signer.sign("hello from mcp")?;
signer.verify(&message)?;

let redactor = CredentialRedactor::new()?;
let result = redactor.redact("Authorization: Bearer super-secret-token");
assert!(result.sanitized.contains("[REDACTED_BEARER_TOKEN]"));
# Ok::<(), agentmesh_mcp::McpError>(())

API Overview

Client (lib.rs)

Unified governance client combining all modules.

Function / Method Description
AgentMeshClient::new(agent_id) Create a client with defaults
AgentMeshClient::with_options(agent_id, opts) Create a client with custom config
client.execute_with_governance(action, context) Run action through governance pipeline

Policy (policy.rs)

YAML-based policy engine with four-way decisions (allow / deny / requires-approval / rate-limit).

Function / Method Description
PolicyEngine::new() Create an empty policy engine
engine.load_from_yaml(yaml) Load rules from a YAML string
engine.load_from_file(path) Load rules from a YAML file
engine.evaluate(action, context) Evaluate an action against loaded policy

Trust (trust.rs)

Integer trust scoring (0–1000) across five tiers with optional JSON persistence.

Function / Method Description
TrustManager::new(config) Create a trust manager
TrustManager::with_defaults() Create with default config
manager.get_trust_score(agent_id) Get current trust score
manager.is_trusted(agent_id) Check against threshold
manager.record_success(agent_id) Increase trust after success
manager.record_failure(agent_id) Decrease trust after failure

Trust tiers:

Tier Score Range
VerifiedPartner 900–1000
Trusted 700–899
Standard 500–699
Probationary 300–499
Untrusted 0–299

Audit (audit.rs)

SHA-256 hash-chained audit log for tamper detection.

Function / Method Description
AuditLogger::new() Create an audit logger
logger.log(agent_id, action, decision) Append an audit entry
logger.verify() Verify chain integrity
logger.get_entries(filter) Query entries by filter

Identity (identity.rs)

Ed25519-based agent identity with DID support.

Function / Method Description
AgentIdentity::generate(agent_id, capabilities) Create a new identity
identity.sign(data) Sign data with private key
identity.verify(data, sig) Verify a signature
identity.to_json() Serialise public identity
AgentIdentity::from_json(json) Deserialise public identity

Policy YAML Format

version: "1.0"
agent: my-agent
policies:
  - name: capability-gate
    type: capability
    allowed_actions:
      - "data.read"
      - "data.write"
    denied_actions:
      - "shell:*"

  - name: deploy-approval
    type: approval
    actions:
      - "deploy.*"
    min_approvals: 2

  - name: api-rate-limit
    type: rate_limit
    actions:
      - "api.call"
    max_calls: 100
    window: "60s"

Execution Rings (rings.rs)

Four-level privilege model inspired by hardware protection rings.

Function / Method Description
RingEnforcer::new() Create a new enforcer with no assignments
enforcer.assign(agent_id, ring) Assign an agent to a ring
enforcer.get_ring(agent_id) Get assigned ring (if any)
enforcer.check_access(agent_id, action) Check if action is permitted
enforcer.set_ring_permissions(ring, actions) Configure allowed actions for a ring

Ring levels:

Ring Level Access
Admin 0 All actions allowed
Standard 1 Configurable actions
Restricted 2 Configurable actions
Sandboxed 3 All actions denied 
use agentmesh::{RingEnforcer, Ring};

let mut enforcer = RingEnforcer::new();
enforcer.set_ring_permissions(Ring::Standard, vec!["data.read".into(), "data.write".into()]);
enforcer.assign("my-agent", Ring::Standard);

assert!(enforcer.check_access("my-agent", "data.read"));
assert!(!enforcer.check_access("my-agent", "shell:rm"));

Agent Lifecycle (lifecycle.rs)

Eight-state lifecycle model tracking an agent from provisioning through decommissioning.

Function / Method Description
LifecycleManager::new(agent_id) Create a new manager (starts in Provisioning)
manager.state() Get current lifecycle state
manager.events() Get recorded transition events
manager.transition(to, reason, initiated_by) Transition to a new state
manager.can_transition(to) Check if a transition is valid
manager.activate(reason) Convenience: transition to Active
manager.suspend(reason) Convenience: transition to Suspended
manager.quarantine(reason) Convenience: transition to Quarantined
manager.decommission(reason) Convenience: transition to Decommissioning

Lifecycle states: Provisioning -> Active <-> Suspended / Rotating / Degraded -> Quarantined -> Decommissioning -> Decommissioned

use agentmesh::{LifecycleManager, LifecycleState};

let mut mgr = LifecycleManager::new("my-agent");
mgr.activate("initial boot").unwrap();
assert_eq!(mgr.state(), LifecycleState::Active);

mgr.suspend("maintenance window").unwrap();
assert_eq!(mgr.state(), LifecycleState::Suspended);

mgr.activate("maintenance complete").unwrap();
assert_eq!(mgr.events().len(), 3);

License

See repository root LICENSE.