The evidence layer for AI agents. Audit trails, approvals, policy gates, and compliance reports for every action.
Website | Docs | SDK Guide | Compliance
Asqav SDK - Python + TypeScript
The official client SDKs for Asqav, the evidence layer for AI agents. Sign every action with ML-DSA-65 (FIPS 204), enforce policies before execution, and produce regulator-ready audit trails.
This repository ships two SDKs from a single monorepo:
python/- the original Python SDK, published to PyPI asasqavtypescript/- the TypeScript SDK, published to npm as@asqav/sdk
Both wrap the same Asqav API. Pick the one that matches your stack.
Install
Python:
pip install asqav
TypeScript / Node.js:
npm install @asqav/sdk
Both packages have zero native dependencies. All ML-DSA cryptography runs server-side.
30-second example
Python:
import asqav
asqav.init(api_key="sk_...")
agent = asqav.Agent.create("my-agent")
sig = agent.sign("api:call", {"model": "gpt-4"})
print(sig.signature_id, sig.verify_url)
TypeScript:
import { init, Agent } from "@asqav/sdk";
init({ apiKey: "sk_..." });
const agent = await Agent.create({ name: "my-agent" });
const sig = await agent.sign({
actionType: "api:call",
context: { model: "gpt-4" },
});
console.log(sig.signatureId, sig.verificationUrl);
Each signed action returns a receipt like:
{
"signature_id": "sig_a1b2c3",
"agent_id": "agt_x7y8z9",
"action": "api:call",
"algorithm": "ML-DSA-65",
"timestamp": "2026-04-27T18:30:00Z",
"chain_hash": "b94d27b9934d3e08...",
"verify_url": "https://asqav.com/verify/sig_a1b2c3"
}
The agent now has a cryptographic identity, a signed audit trail, and a verifiable action record.
Data handling modes
The SDK auto-detects whether you're pointing at the Asqav cloud or a self-hosted deployment, and selects the safer default for each:
- *Cloud (`.asqav.com`)**: hash-only by default. The SDK builds a fingerprint of your action context, computes a SHA-256 hash locally, and sends only the hash plus a small metadata bag of action_type, agent_id, session_id, model_name, and tool_name. Raw prompts and tool arguments stay on your side.
- Self-hosted: full-payload by default. The server can run policy checks, PII redaction, and richer audit. Recommended when you control the deployment.
Override anytime:
# Python
asqav.init(api_key="...", base_url="https://api.asqav.com", mode="hash-only")
// TypeScript
await init({ apiKey: "...", baseUrl: "https://api.asqav.com", mode: "hash-only" });
The fingerprint format is RFC 8785-style sorted JSON with no whitespace, hashed with SHA-256. See docs/fingerprint-spec.md and conformance/vectors.json for the spec and cross-language test vectors.
For self-hosted and split-trust deployments, bring-your-own KMS, SCITT and COSE receipt export, and air-gapped mode, see asqav.com/docs.
Standards
Asqav's compliance receipts are profiled in IETF Internet-Draft draft-marques-asqav-compliance-receipts, an Independent Submission that profiles draft-farley-acta-signed-receipts for EU AI Act Articles 12 and 26, and DORA Article 17 bindings.
Coding agents
Autonomous coding agents open and merge their own pull requests, and a git author string is mutable. The protectmcp:lifecycle:code_authorship receipt binds a change to the agent that authored it: the repository, the commit and base SHA, a digest of the diff, the change class, and a producer-asserted authored_by block, signed with ML-DSA-65 and chained per agent. Asqav records that the change existed, was key-authored, and was chained at a point in time. It does not clone the repository, re-run the diff, or judge whether the code is correct. A ready-to-use GitHub Action lives in github-action/ and can also emit the receipt as an in-toto attestation. See the code-authorship receipts docs.
Why governance
Without governance, there is no record of what agents did, any agent can do anything, one person approves everything by hand, compliance reports are written manually, and the reasoning is lost once the run ends. Asqav changes each of those:
- Every action is signed with ML-DSA (FIPS 204).
- Policies block dangerous actions before they run.
- Critical actions can require multi-party authorization.
- EU AI Act and DORA reports are generated automatically.
- The prompt, trace, and output are signed and replayable.
Per-language docs
The root README only covers the cross-language story. For language-specific guides - decorators, async clients, framework integrations, CLI commands, local mode, batched signing, three-phase signing, scope tokens, replay, attestations, and more - see:
- Python:
python/README.md - TypeScript:
typescript/README.md
The full reference lives at asqav.com/docs.
What's in each SDK
Both SDKs cover the same core surface: agent identity, signed actions, batched signing, policy enforcement, scope tokens, replay, attestations, three-phase signing, and the user_intent envelope on agent.sign(...), signed with Ed25519, ECDSA P-256, or WebAuthn.
The Python SDK additionally ships:
- The
asqavCLI:quickstart,demo,verify,doctor,agents,sync, plusreplay,preflight,approve,budgetwithcheckandrecord, andcompliancewithframeworksandexport. See asqav.com/docs/cli. - A pytest plugin, enabled with
pytest --asqav, that signs every test result and emits a Merkle-rooted compliance bundle on session finish. See asqav.com/docs/pytest-plugin. - Native callbacks for LangChain, CrewAI, LiteLLM, Haystack, OpenAI Agents SDK, LlamaIndex, smolagents, DSPy, PydanticAI, Letta, Strands, and Instructor through the Hooks API.
- Cookbooks for Streamlit dashboards and Dify workflows under
python/examples/. - Local-mode offline signing with deferred sync.
The TypeScript SDK focuses on the core API and Agent surface for Node 20+ runtimes. It additionally ships:
- The
user_intentenvelope onagent.sign(...). - A Vercel AI SDK adapter at
@asqav/sdk/extras/vercel-aithat plugs intoexperimental_telemetry: { tracer }and signs every span the AI SDK opens. - A LangChain.js callback handler
AsqavCallbackHandlerat@asqav/sdk/extras/langchainthat signs each chain step. - A Mastra hook
AsqavMastraHookat@asqav/sdk/extras/mastrathat signs each step the Mastra agent emits. - An OpenAI Agents JS adapter
AsqavOpenAIAgentsAdapterat@asqav/sdk/extras/openai-agentsthat signs tool calls on the agent runner.
If a feature exists in one SDK but not the other, the language README will mark it explicitly.
Repository layout
asqav-sdk/
python/ Python SDK (pip install asqav)
src/
tests/
pyproject.toml
README.md
typescript/ TypeScript SDK (npm install @asqav/sdk)
src/
tests/
package.json
README.md
conformance/ Cross-language conformance fixtures both SDKs run against
.github/workflows/
ci.yml Path-filtered CI for both languages
publish.yml Tag-based publish (py-v* to PyPI, ts-v* to npm)
CHANGELOG.md
README.md (this file)
The two SDKs are versioned and released independently.
Verifying a receipt
Receipts are portable. Anyone with a signature_id can verify it without an API key:
Python:
import asqav
result = asqav.verify("sig_a1b2c3")
assert result["verified"]
print(result["agent_id"], result["chain_hash"])
TypeScript:
import { verify } from "@asqav/sdk";
const result = await verify("sig_a1b2c3");
console.assert(result.verified);
console.log(result.agentId, result.chainHash);
Or open the receipt's verify_url in a browser. Hashes are reproducible offline from the RFC 8785 payload, the JSON canonicalization format, so auditors do not need to trust Asqav's servers - the signature speaks for itself.
Counterparty acknowledgment
When two agents hand off a workflow, where A signs an action and B acts on it, the SDK lets B emit a protectmcp:acknowledgment receipt that cryptographically binds B's bytes to A's. The bundle carries A's full envelope, B's acknowledgment, and a SHA-256 binding the verifier recomputes offline. A tampered intermediary cannot mutate the bytes without breaking the equality, so a regulator can verify the handoff with one digest comparison.
Python:
from asqav import compute_counterparty_binding, verify_counterparty_binding
binding = compute_counterparty_binding(originating_envelope)
b_payload["counterparty_binding"] = binding.to_wire()
# ... B signs b_payload normally ...
outcome = verify_counterparty_binding(b_envelope, originating_envelope)
assert outcome.valid, outcome.label
TypeScript:
import { computeCounterpartyBinding, verifyCounterpartyBinding } from "@asqav/sdk";
const binding = computeCounterpartyBinding(originatingEnvelope);
bPayload.counterparty_binding = binding;
// ... B signs bPayload normally ...
const outcome = verifyCounterpartyBinding(bEnvelope, originatingEnvelope);
if (!outcome.valid) throw new Error(outcome.label);
The cloud /verify endpoint reports the same outcome on the verification response under counterparty_binding_verified, and /audit-pack/export pulls the originating receipt into the bundle automatically when an acknowledgment in the export window references one outside it.
Conformance
The conformance/ directory contains shared test fixtures both SDKs run against. Adding a feature means adding a fixture there first, then making both SDKs pass it. CI reruns both matrices whenever conformance/ changes, even if neither python/ nor typescript/ was touched, so cross-language drift is caught at PR time.
CI
CI runs on every push to main and every pull request. It tests only the language whose source changed: pytest on Python 3.10, 3.11, and 3.12, and npm test on Node 20 and 22. A change to conformance/ runs both. The aggregator job ci-ok is the single required status check for branch protection.
Run governance checks in your own CI
asqav doctor validates configuration and connectivity in any environment with ASQAV_API_KEY set, returning non-zero on failure so it gates PRs cleanly. Pair it with asqav.compliance.export_bundle to package signed receipts from your test runs into a self-contained, Merkle-rooted artifact for auditors. See docs/github-actions.md for a copy-paste workflow.
Contributing
See CONTRIBUTING.md for the full contributor guide. The short version:
- Fork the repo and create a feature branch.
- Run the relevant test suite locally.
- Python:
cd python && pip install -e ".[all]" pytest && pytest tests -v - TypeScript:
cd typescript && npm ci && npm test
- Python:
- Open a pull request against
main. CI must go green for the PR to land.
Direct pushes to main are blocked. Every change lands through a PR.
Releases
Both SDKs ship to different registries on independent cadences, driven by prefixed git tags: py-v* publishes the Python half to PyPI, ts-v* publishes the TypeScript half to npm. Bump the version in the relevant manifest, update CHANGELOG.md, then tag and push. The publish.yml workflow inspects the tag prefix and runs only the matching publish job. Python uses PyPI's OIDC trusted publisher. TypeScript publishes via the NPM_TOKEN repo secret.
Ecosystem
asqav(PyPI): the Python SDK.@asqav/sdk(npm): the TypeScript SDK.- asqav-mcp: MCP server for Claude Desktop, Claude Code, and Cursor.
- asqav-compliance: CI/CD compliance scanner.
Plans
Asqav has two plans: Free and Enterprise.
Free covers 5,000 signatures/month, 10 agents, 10 policies, 3 team members, and 30-day log retention. Most features are included: OpenTimestamps anchoring, the MCP server, audit export, framework integrations, content scanning, OpenTelemetry export, the Replay API, approvals, governance query and attestation, and 2 compliance reports/month.
Enterprise adds RFC 3161 timestamps, SSO/SAML, managed and bring-your-own KMS, SD-JWT selective disclosure, IP allowlists, multi-party quorum signing, quarantine and incident management, a self-hosted signer, and dedicated support. See asqav.com/pricing for the full breakdown.
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
- Docs: https://asqav.com/docs
- Repo: https://github.com/jagmarques/asqav-sdk
- Blog: https://dev.to/jagmarques
- Dashboard: https://asqav.com/dashboard
License
MIT - see LICENSE for details.
If Asqav helps you, consider giving the repo a star. It helps others find the project.