asqav 0.2.14

AI agent governance - audit trails, policy enforcement, compliance

Governance for AI agents. Audit trails, policy enforcement, and compliance.

Website | Docs | 中文文档 | SDK Guide | Compliance

asqav SDK

Thin Python SDK for asqav.com. All ML-DSA cryptography runs server-side. Zero native dependencies.

Install

pip install asqav

import asqav

asqav.init(api_key="sk_...")
agent = asqav.Agent.create("my-agent")
sig = agent.sign("api:call", {"model": "gpt-4"})

Your agent now has a cryptographic identity, a signed audit trail, and a verifiable action record.

Each signed action returns:

{
  "signature_id": "sig_a1b2c3",
  "agent_id": "agt_x7y8z9",
  "action": "api:call",
  "algorithm": "ML-DSA-65",
  "timestamp": "2026-04-06T18:30:00Z",
  "chain_hash": "b94d27b9934d3e08...",
  "verify_url": "https://asqav.com/verify/sig_a1b2c3"
}

Why

Without governance	With asqav
No record of what agents did	Every action signed with ML-DSA (FIPS 204)
Any agent can do anything	Policies block dangerous actions in real-time
One person approves everything	Multi-party authorization for critical actions
Manual compliance reports	Automated EU AI Act and DORA reports
Breaks when quantum computers arrive	Quantum-safe from day one

Decorators and context managers

@asqav.sign
def call_model(prompt: str):
    return openai.chat.completions.create(model="gpt-4", messages=[{"role": "user", "content": prompt}])

with asqav.session() as s:
    s.sign("step:fetch", {"source": "api"})
    s.sign("step:process", {"records": 150})

Async support

agent = await asqav.AsyncAgent.create("my-agent")
sig = await agent.sign("api:call", {"model": "gpt-4"})

All API calls retry automatically with exponential backoff on transient failures.

CLI

pip install asqav[cli]

asqav quickstart              # one-command project setup
asqav doctor                  # health check for your configuration
asqav verify sig_abc123
asqav agents list
asqav agents create my-agent
asqav sync

Local mode

Sign actions offline when the API is unreachable. Queue syncs when connectivity returns.

from asqav import local_sign

local_sign("agt_xxx", "task:complete", {"result": "done"})
# Later: asqav sync

Works with your stack

Native integrations for 9 frameworks. Each extends AsqavAdapter for version-resilient signing.

pip install asqav[langchain]
pip install asqav[crewai]
pip install asqav[litellm]
pip install asqav[haystack]
pip install asqav[openai-agents]
pip install asqav[llamaindex]
pip install asqav[smolagents]
pip install asqav[dspy]
pip install asqav-pydantic          # PydanticAI (external package)
pip install asqav[all]

LangChain

from asqav.extras.langchain import AsqavCallbackHandler

handler = AsqavCallbackHandler(api_key="sk_...")
chain.invoke(input, config={"callbacks": [handler]})

# Observe mode - test governance without API calls
handler = AsqavCallbackHandler(api_key="sk_...", observe=True)

CrewAI

from asqav.extras.crewai import AsqavCrewHook

hook = AsqavCrewHook(api_key="sk_...")
task = Task(description="Research competitors", callbacks=[hook.task_callback])

LlamaIndex

from asqav.extras.llamaindex import AsqavLlamaIndexHandler

handler = AsqavLlamaIndexHandler(api_key="sk_...")

LiteLLM / Haystack / OpenAI Agents SDK / smolagents / DSPy

from asqav.extras.litellm import AsqavGuardrail
from asqav.extras.haystack import AsqavComponent
from asqav.extras.openai_agents import AsqavGuardrail
from asqav.extras.smolagents import AsqavSmolagentsHook
from asqav.extras.dspy import AsqavDSPyCallback

PydanticAI

pip install asqav-pydantic

from asqav_pydantic import AsqavHooks

See integration docs for full setup guides.

Enforcement

asqav provides three tiers of enforcement for AI agent governance:

Strong - the asqav-mcp server acts as a non-bypassable tool proxy. Agents call tools through the MCP server, which checks policy and signs the decision before allowing execution. With a tool_endpoint configured, the server forwards the call and signs request + response together as a bilateral receipt.

Bounded - pre-execution gates (gate_action) check policy and sign the decision before the agent acts. After completing the action, the agent calls complete_action to close the bilateral receipt, linking the approval to the outcome.

Detectable - every action gets a quantum-safe signature (ML-DSA-65) hash-chained to the previous one. If logs are tampered with or entries omitted, the chain breaks and verification fails.

Most teams use all three tiers for different tools. High-risk mutations go through the strong path, routine operations use bounded, and everything gets the detectable layer. Bilateral receipts ensure auditors can verify both what was authorized and what actually happened.

Policies

asqav.create_policy(
    name="no-deletions",
    action_pattern="data:delete:*",
    action="block_and_alert",
    severity="critical"
)

MCP enforcement

pip install asqav-mcp
export ASQAV_PROXY_TOOLS='{"sql:execute": {"risk_level": "high", "require_approval": true}}'
asqav-mcp

The MCP server exposes gate_action (bounded) and enforced_tool_call (strong) tools to any MCP client - Claude Desktop, Claude Code, Cursor, or custom agents.

Multi-party signing

Distributed approval where no single entity can authorize alone:

config = asqav.create_signing_group("agt_xxx", min_approvals=2, total_shares=3)
session = asqav.request_action("agt_xxx", "finance.transfer", {"amount": 50000})
asqav.approve_action(session.session_id, "ent_xxx")

Semantic action patterns

Use descriptive labels instead of glob patterns for clearer policies:

sig = agent.sign("sql-destructive", {"query": "DROP TABLE users"})

Pre-flight checks

Combine revocation, suspension, and policy checks in a single call before running sensitive actions. Returns human-readable explanations of what each policy would do:

result = agent.preflight("data:delete")
if not result.cleared:
    raise RuntimeError(f"blocked: {result.reasons}")
# result.explanations contains plain-English descriptions of each check

Reproducibility metadata

Attach system prompt hashes and other reproducibility context to signatures:

sig = agent.sign("llm:call", ctx, system_prompt_hash="sha256:a1b2c3...")

Compliance bundle export

Export signed action trails as regulation-specific compliance bundles:

bundle = asqav.export_bundle(signatures, "eu_ai_act_art12")

Scope tokens

Portable, short-lived tokens that let a receiving service verify what an agent is allowed to do without calling your org:

token = agent.create_scope_token(
    actions=["data:read:customer"],
    ttl=3600
)
# Attach to outgoing API calls
requests.get(url, headers=token.to_header())
# Receiving service verifies without calling your org
verified = asqav.verify_scope_token(token_string)

Replay

Reconstruct what an agent did in a session, with chain integrity verification:

timeline = asqav.replay(agent_id="agt_abc", session_id="sess_xyz")
print(timeline.summary())
assert timeline.chain_integrity  # verify no tampering

# Or replay from offline bundle
timeline = asqav.replay_from_bundle(bundle)

Output verification

Bind a tool output to its input and verify later that nothing was tampered with:

import hashlib, json

prompt = {"query": "top customers"}
input_hash = hashlib.sha256(json.dumps(prompt, sort_keys=True).encode()).hexdigest()

output = run_tool(prompt)
sig = agent.sign_output("tool:search", input_hash, output)

# Later, from any verifier
result = asqav.verify_output(sig.signature_id, output)
assert result["verified"]

Budget tracking

Client-side spend tracker that persists every cost as a signed record. Fails closed when the limit would be exceeded:

budget = asqav.BudgetTracker(agent, limit=10.0, currency="USD")

decision = budget.check(estimated_cost=0.25)
if not decision.allowed:
    raise RuntimeError(decision.reason)

budget.record("api:openai", actual_cost=0.23, context={"model": "gpt-4"})

Attestations

Generate a portable, signed attestation document proving an agent's governance status. Share it with auditors or partners for independent verification:

attestation = asqav.generate_attestation("agt_abc123", session_id="sess_xyz")

# Anyone with the document can verify it
result = asqav.verify_attestation(attestation)
assert result["valid"] and result["all_valid"]

Features

• Signed actions - every agent action gets a ML-DSA-65 signature with RFC 3161 timestamp
• Decorators - @asqav.sign wraps any function with cryptographic signing
• Async - full async support with AsyncAgent and automatic retry
• CLI - verify signatures, manage agents, quickstart setup, doctor health check
• Local mode - sign actions offline, sync later
• Observe mode - test governance without API calls using observe=True
• Framework integrations - LangChain, CrewAI, LiteLLM, Haystack, OpenAI Agents SDK, LlamaIndex, smolagents, DSPy, PydanticAI
• Semantic action patterns - descriptive labels instead of glob patterns
• Reproducibility metadata - attach system prompt hashes to signatures
• Compliance bundles - export regulation-specific bundles with export_bundle
• Three-tier enforcement - strong (tool proxy), bounded (pre-execution gates), detectable (signed audit trail)
• Policy enforcement - block or alert on action patterns before execution
• Multi-party signing - m-of-n approval using threshold ML-DSA
• Pre-flight checks - combined revocation, suspension, and policy gate in one call
• Output verification - bind tool outputs to inputs and verify later with sign_output / verify_output
• Budget tracking - client-side spend limits with signed, replayable records
• Attestations - portable signed documents proving an agent's governance status
• Agent identity - create, suspend, revoke, and rotate agent keys
• Audit export - JSON/CSV trails for compliance reporting
• Tokens - scoped JWTs and selective-disclosure tokens (SD-JWT)
• Scope tokens - portable, short-lived tokens for cross-org action verification
• Replay - reconstruct agent sessions with chain integrity checks

In the press

• Help Net Security: Asqav - Open-source SDK for AI agent governance

Ecosystem

Package	What it does
asqav	Python SDK
asqav-mcp	MCP server for Claude Desktop/Code
asqav-compliance	CI/CD compliance scanner

Free tier

Get started at no cost. Free tier includes agent creation, signed actions, three-tier enforcement, policies, audit export, and framework integrations. Threat detection and monitoring on Pro ($39/mo). Compliance reports, multi-party signing, and incident management on Business ($149/mo). See asqav.com for pricing.

Discovery

Machine-readable service descriptor at https://asqav.com/.well-known/governance.json for external tools that want to auto-discover asqav's endpoints, algorithms, capabilities, and integrations.

Links

• Integration Docs
• Blog
• Dashboard
• PyPI

License

MIT - see LICENSE for details.

---

If asqav helps you, consider giving it a star. It helps others find the project.