MandateGuard

Deterministic, auditable payment policy for autonomous AI agents.

MandateGuard is a pre-action enforcement layer that sits between an agent and its tools/wallet. Every tool call that moves money is evaluated by a pure, deterministic engine — budgets, allowlists, denylists, rate limits, and signed mandates — before anything executes. No LLM is ever in the decision path, which is exactly what makes every verdict reproducible and every ledger entry verifiable.

It ships with an MCP server so any agent (Claude, Cursor, or your own harness) can mount it as a guardrail in minutes.

Why

The 2026 agentic-economy reality:

• OWASP LLM08 — Excessive Agency is one of the top LLM app risks. Agents given wallets are getting drained: the SoK on agentic commerce documents $40M+ in real losses (drain attacks, memory poisoning, tool abuse).

• Payment standards (Google AP2's Intent/Cart/Payment mandates, Coinbase x402, ERC-8004) define what a mandate is — but none of them ship the enforcement layer that actually blocks an agent mid-flight.

• Gartner: 40% of enterprise apps will embed agents by end of 2026. Those agents will move money. They need rails.

The market gap: a deterministic (non-LLM) policy engine + audit trail + MCP distribution. That is this repo.

Features

• Deterministic engine — same inputs, same verdict, always. Auditable by replay, no model sampling in the decision.

• Per-actor scopes — allowed tools, allowed destinations, per-call max, currency, per-window call limits.

• Global guards — total budget caps, destination allowlist/denylist.

• Signed mandates (Ed25519) — short-lived, nonce-bound, issuer-signed authorizations in the AP2 / x402 style. An agent cannot widen its own scope.

• Tamper-evident ledger — every decision is append-only and SHA-256 chained. Any edit, reorder, or deletion is detected by a linear scan.

• MCP server — mount as a guardrail; tools for policy, authorize, mandate issuance, and ledger health.

• Zero deps in the decision path — cryptography only for mandates; the core rules run on the stdlib alone.

Install

# from this repo (works today; also on the official MCP Registry)
git clone https://github.com/ezequiellich44-cmd/MandateGuard.git
cd MandateGuard
python -m pip install -e .

# or directly from the source:
python -m pip install "git+https://github.com/ezequiellich44-cmd/MandateGuard.git"

Note: mandateguard on PyPI is pending Trusted Publisher setup; until then the repo URL is the canonical install path. The MCP bundle is live on the official MCP Registry (io.github.ezequiellich44-cmd/mandateguard), so MCP-aware clients can install it without any Python step.

Quickstart

from mandateguard import Intent, Policy, PolicyEngine, Scope

policy = Policy(
    scopes={
        "wallet-agent": Scope(
            tools=("pay",),
            destinations=("0xGOOD",),
            max_amount=1000,          # per call
            currency="usd",
            max_calls_per_window=5,
        )
    },
    global_max_amount=2000,           # per actor
    allowlist=("0xGOOD",),
    denylist=("0xSCAM",),
)
engine = PolicyEngine(policy)

decision = engine.authorize(
    Intent(tool="pay", destination="0xGOOD", amount=800, actor="wallet-agent")
)
print(decision.status)   # DecisionStatus.APPROVED

Denied calls are blocked with structured reasons; state (spend/rate) commits only on approval, so replays are deterministic.

MCP server

The package ships an installable MCP server entrypoint:

python -m pip install -e ".[mcp]"
mandateguard-mcp            # stdio server, ready for Claude/Cursor/harness

For Claude Code:

claude mcp add mandateguard -- mandateguard-mcp

MandateGuard is published on the official MCP Registry: io.github.ezequiellich44-cmd/mandateguard (version 1.0.0, mcpb bundle, active). MCP-aware clients that sync the registry can discover and install it directly. The bundle ships the same stdio server and 14-tool surface.

Exposed tools: set_scope, set_global_policy, authorize, init_ledger, ledger_status, create_mandate_signer, issue_mandate, check_mandate, activate_license, license_status, reset_state, plus Pro-gated revoke_mandate and persist_state behind a signed Pro license (USDT purchase — see the Buy section).

Mandates

from mandateguard import Mandate, MandateSigner, verify_mandate

issuer = MandateSigner()
m = Mandate(actor="wallet-agent", max_amount=500, currency="usd",
            tools=("pay",), destinations=("0xGOOD",),
            not_before="2026-01-01T00:00:00+00:00",
            not_after="2099-01-01T00:00:00+00:00", nonce="abc", issuer="you")
sig = issuer.sign(m)
verify_mandate(issuer.public_key_bytes, m, sig)   # True

Architecture

See docs/ARCHITECTURE.md for the decision flow and state model, docs/THREAT_MODEL.md for what this does and does not protect against, and docs/LAUNCH.md for the commercial pitch and go-to-market kit.

Agent intent ──> authorize(intent) ──> PolicyEngine
                                          │  scope? allowlist? denylist?
                                          │  budget? rate limit? mandate?
                                          ▼
                                     APPROVED / DENIED / REQUIRES_APPROVAL
                                          │
                                          ▼
                              append-only SHA-256 ledger (audit)

Tests

python -m pytest -q

License

MIT. See LICENSE.