Skip to main content

Mandata

OpenSSF Scorecard

The mandate toolkit for agent-initiated payments.

When AI agents hold spending authority, someone must answer: what exactly are they allowed to do — and can you prove it afterward? Mandata is a reference implementation of the bounded-autonomy mandate chain:

principal → agent → credential → transaction → evidence

The problem

Agentic payment protocols (AP2, x402, card-network agent programs) let AI agents initiate real financial transactions. But production systems lack a standard way to:

  • Express mandates — bounded permissions: amounts, payees, time windows, velocity limits
  • Bind credentials — cryptographic proof a transaction was made within mandate
  • Enforce at the gate — deterministic checks before settlement, not logs after
  • Produce evidence — an auditable chain showing every agent action against its mandate

Mandata formalizes all four as a small, dependency-light library.

Quickstart

pip install mandata
from mandata import Mandate, Bounds, enforce, claim, settle, to_minor

# A principal grants a bounded mandate to an agent (amounts in minor units)
mandate = Mandate(
    principal="user:alice",
    agent="agent:shopping-bot",
    bounds=Bounds(
        max_amount=to_minor("500.00", "GBP"),  # 50000 pence per transaction
        currency="GBP",
        allowed_payees=["store.example"],
        expires_in_days=30,
        velocity_limit=3,  # max authorizations per day
    ),
)

# At transaction time: deterministic enforcement BEFORE settlement.
# tx_id makes retries idempotent and links the authorization to settlement.
decision = enforce(mandate, payee="store.example", amount=4999, currency="GBP", tx_id="order-1")
assert decision.allowed  # → True (status AUTHORIZED, amount reserved)
print(decision.check)  # → "WITHIN" (normative MMF check ID)

# Claim takes the single dispatch slot and verifies the rail instruction binds
# to the approval — only then submit, and only once.
claim(mandate, "order-1", "store.example", 4999, "GBP")
...  # rail.pay(..., idempotency_key="order-1")
settle(mandate, "order-1")  # or void(mandate, "order-1") if it failed

Architecture

principal issues mandate ──► credential binds mandate to agent
        │                           │
        ▼                           ▼
  agent proposes transaction ──► mandate gate (enforce())
        │                           │
        ▼                           ▼
   allowed? ──► yes: execute    no: reject
        │                           │
        └───────► evidence record appended (hash-chained audit log)

See docs/architecture.md for the full model, docs/mandate-model.md for the schema, and docs/diagrams.md for the system design diagrams.

Roadmap

  • v0.1 — Mandate model, bounds enforcement, evidence chain (Python)
  • Credential signing (Ed25519) + verification API
  • Revocation registry (mandate kill-switch at transaction speed)
  • AP2 / x402 / LangChain adapter layer
  • Multi-hop delegation (agent → sub-agent mandate weakening)
  • v0.3 — MMF-v0.2: credential-bound gate, chain-wide delegation enforcement, fail-closed registry, canonical evidence hashing, 43 conformance vectors
  • v0.4 — MMF-v0.3: integer minor-unit amounts, tx_id idempotency, authorize/settle/void lifecycle, payee binding in adapters, gate-signed evidence, cross-process serializable SQLite store, 58 conformance vectors
  • v0.4.1 — MMF-v0.3.1: mandatory transaction currency, INVALID_CHAIN for unresolvable/cyclic chains, chain-wide credential verification, claim()/unclaim() single-dispatch lifecycle (AUTHORIZED→CLAIMED→SETTLED), atomic durable evidence in the state transaction, byte-level canonical serialization, 64 conformance vectors
  • v0.4.2 — MMF-v0.3.2: permanently-consumed claim slot (no re-claim; recovery = void + new tx_id), REJECTED evidence for failed lifecycle transitions, velocity counts issued authorizations (void refunds budget only), budget-window attenuation, canonical evidence v2 (currency + status hashed, NFC), 65 conformance vectors
  • v0.4.3 — MMF-v0.3.3: void of a CLAIMED transaction requires confirmed_rail_cancelled (uncertain in-flight stays CLAIMED), serialized ledger-first ordering + TX_CONFLICT evidence-only rule, mediation-scoped claim guarantee, REJECTED records carry the attempted instruction, byte-level timestamp rule, per-gate evidence chain scope — 66 vectors
  • v0.4.4 — MMF-v0.3.4: canonical evidence v3 — length-prefixed fields (any value encodable), integer-millisecond timestamps, hashed REFERENCE provenance; void of a CLAIMED transaction requires a cancellation reference; checked arithmetic on accumulated totals
  • v0.4.5 — MMF-v0.3.5: three-level dispatch guarantee formalized (claim / trusted dispatcher / rail), normative evidence serialization table (types, nulls, NFC, encodings), evidence obligations scoped to completed durably-committed evaluations
  • v0.4.6 — MMF-v0.3.6: revocation re-verified at claim (a revoked mandate's unclaimed authorization cannot be dispatched); per-state revocation policy; principal-authorized delegation model stated; exact unnormalized request equality
  • v0.4.7 — MMF-v0.3.7: claim re-verifies revocation AND expiry chain-wide (linearization = claim commit); unresolved open reservations consume budget at any age (L17/L18); optional signed-evidence conformance profile; typed-value request equality — 68 vectors
  • v0.4.17 — MMF-v0.3.17: cross-mandate TX_CONFLICT evidence; rail idempotency-key rule; normative window arithmetic (inclusive lower bound) — vectors L27, L28, D18 — 85 vectors
  • v0.4.16 — MMF-v0.3.16: cross-mandate evidence-leak fix; rejection causes/precedence now conformance-tested (L21–L26) — 82 vectors; attenuation proof scoped to one hypothetical history
  • v0.4.15 — MMF-v0.3.15: structural vs effective attenuation formalized (+ randomized test); settle wording; atomic-commit rule; evidence forensic limit — no behavior change
  • v0.4.14 — MMF-v0.3.14: replay-first ordering structural (resolution only for new tx_id); idempotent lifecycle step precedes STATUS; vector L20 — 76 vectors
  • v0.4.13 — MMF-v0.3.13: final semantic clarifications — binding scope, repeated-lifecycle-call evidence rule, check terminology; no behavior change
  • v0.4.12 — MMF-v0.3.12: claim-time revocation guarantee scoped to visible state; settle-after-window aging rule + vector L19 — 75 vectors
  • v0.4.11 — MMF-v0.3.11: settle requires CLAIMED (AUTHORIZED can only claim or void); per-operation REJECTED precedence with REVOKED_REGISTRY; evidence-fidelity qualification
  • v0.4.10 — MMF-v0.3.10: resolution-failure-as-result ordering; fixed REJECTED cause precedence; exact-bytes evidence (NFC removed); negative credential vectors C02–C06 — 74 vectors
  • v0.4.9 — MMF-v0.3.9: ledger-lookup-first ordering made structural (chain re-resolved under serialization); normative <OP>:<CAUSE> on REJECTED evidence; cancellation-reference binding rules; credential golden vector C01 — 69 vectors
  • v0.4.8 — MMF-v0.3.8: unified length-prefixed canonicalization for credentials + bounds_hash (delimiter-smuggling closed); mandate_id immutable-for-life + durable mandate preservation; timestamp edge cases; typed-equality integral rule
  • TypeScript implementation (core gate scaffold only — not yet MMF-v0.3.8 conformant)

Research

This project is the open-source implementation arm of the survey:

Agentic AI in Payments and Financial Transactions: A Survey of Architectures, Safeguards and Open Problems — S.C. Aji, 2026. doi:10.13140/RG.2.2.13674.89288

Contributing

We welcome contributors — see CONTRIBUTING.md. Good entry points are labelled good first issue.

License

Apache-2.0 — see LICENSE.

Demo app

examples/demo_app.py runs a full simulated deployment over real HTTP — a revocation-registry service, a mock payment rail, and an agent whose every proposed payment goes through the gate: signed credentials, sub-agent delegation, x402 intake, compromise attempts, remote revocation, then a full evidence-chain audit. python examples/demo_app.py — stdlib only.

For a shorter single-process walkthrough see examples/end_to_end.py.

Metadata

Release files for mandata 0.4.17

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for mandata 0.4.17
File Size Uploaded Provenance Actions
mandata-0.4.17.tar.gz 51.3 kB With provenance Details for mandata-0.4.17.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for mandata 0.4.17
File Interpreter ABI Platform Actions
mandata-0.4.17-py3-none-any.whl Python 3 none any Details for mandata-0.4.17-py3-none-any.whl

Total release size: 90.9 kB

Release files / mandata-0.4.17.tar.gz

Download URL mandata-0.4.17.tar.gz
Size 51.3 kB
Tags Source
SHA-256 checksum
How to use checksums
a50608974e502b59d763628f08126d378bb68f7070ea6266d15597b4ef42bfcd
BLAKE2b-256 checksum
How to use checksums
94ea952eb333ae839c639f9064b288addc958bfcd1ace63c93b74a482ba31005
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

Release files / mandata-0.4.17-py3-none-any.whl

Download URL mandata-0.4.17-py3-none-any.whl
Size 39.6 kB
Tags Python 3
SHA-256 checksum
How to use checksums
4e9cb768b9be49a268741363e107ba9048acd698ca6219a68076659fb6c58ec1
BLAKE2b-256 checksum
How to use checksums
08d236a577f8313e4cd7f7d0080ad6c8ab4df2b3accb3380aa16bed812eaacc4
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

Security for mandata 0.4.17

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for mandata-0.4.17.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page