Skip to content

README_PYPI.md

Latest commit

 

History

History
116 lines (84 loc) · 4.58 KB

README_PYPI.md

File metadata and controls

116 lines (84 loc) · 4.58 KB

AgentVeil

Action control for autonomous agents — check posture, gate risky actions, prove execution.

AgentVeil is the Python SDK for agent action control: posture checks, Runtime Gate decisions, signed receipts, W3C verifiable credentials, plus DID identity, reputation signals, and MCP integrations.

pip install agentveil

Quick Start

Run locally with real cryptography and mocked HTTP. No server is required.

from datetime import timedelta
from agentveil import AVPAgent

owner = AVPAgent.create(mock=True, name="workflow-owner")
agent = AVPAgent.create(mock=True, name="demo-agent")
agent.register(display_name="Demo Agent")

delegation = owner.issue_delegation_receipt(
    agent_did=agent.did,
    allowed_categories=["deploy"],
    valid_for=timedelta(minutes=15),
)
verification = agent.verify_delegation_receipt(delegation)

print("delegation valid:", verification["valid"])
print("scope:", verification["scope"][0]["value"])

For production setup, see the Customer Integration guide.

What AgentVeil Provides

  • Posture checks before risky agent actions reach production.
  • Runtime Gate decisions for allow, approval required, or block outcomes.
  • Signed receipts and proof packets for audit and offline verification.
  • W3C VC v2.0 credentials with eddsa-jcs-2022 Data Integrity proofs.
  • DID identity with portable did:key Ed25519 keys.
  • Framework integrations for CrewAI, LangGraph, AutoGen, OpenAI, Claude MCP, Gemini, PydanticAI, Paperclip, and AWS Bedrock.
  • MCP transport proxy for IDE clients (Claude Desktop, Cursor, Cline, Windsurf, VS Code) - wrap downstream MCP servers with Action Control Plane gating via the agentveil-mcp-proxy console script.

AgentVeil makes agent actions constrained, auditable, and reversible within a declared action vocabulary and policy subset. It does not claim to solve the general access-control safety problem; it produces bounded decisions and signed evidence that operators can review.

Offline Verification

Fetch a W3C Verifiable Credential:

curl https://agentveil.dev/v1/reputation/{agent_did}/credential?format=w3c

Verify it with any VC library, or with the SDK:

cred = agent.get_reputation_credential(format="w3c")
assert AVPAgent.verify_w3c_credential(cred)

MCP Server

The base install includes the MCP runtime dependency:

pip install agentveil
agentveil-mcp

Local/full MCP mode exposes Runtime Gate evaluation, human approval routing, approved execution, signed receipt retrieval, reputation checks, identity lookup, and audit verification. Hosted read-only mode exposes public inspection tools only.

The compatibility extra agentveil[mcp] still works for legacy setups. MCP setup details are in the MCP README.

MCP Transport Proxy

The agentveil-mcp-proxy console script wraps a downstream MCP server with runtime decision gating, human approval routing, durable signed evidence, and replay defense. Point your IDE at agentveil-mcp-proxy instead of directly at the downstream server; the proxy applies AVP policy before forwarding.

agentveil-mcp-proxy init
agentveil-mcp-proxy doctor
agentveil-mcp-proxy run

AVP approvals are capability tokens, not flat permissions. They are signed, scoped to action context and payload hash, time-bounded by expiry, guarded against replay at the proxy boundary, and attenuated when follow-on grants such as similar_5m narrow the original approval scope.

See the MCP Proxy README for the full quick start and IDE configuration examples.

Resources

License

MIT. See the license.