io.github.dingdawg/agent-spend-policy-mcp — Model Context Protocol (MCP) Server

This local MCP server performs deterministic checks for agent spend-policy decisions. It evaluates a proposed agent spend action against a supplied policy and returns a decision with a stable reason code, without holding funds, keys, payment credentials, or making network requests.

🛠️ Key Features

Deterministic, local evaluation of proposed agent spend actions
Returns one of ELIGIBLE, DENY, or STEP_UP
Stable reason code for each result
No network access, no network requests
No custody, signing, sending, settling, or recording of payments

🚀 Use Cases

Local governance checks for agent spend eligibility
Policy enforcement logic for AI-agent spend actions
Decisioning workflows that require deterministic policy outcomes

⚡ Developer Benefits

Stable outputs (ELIGIBLE / DENY / STEP_UP) with stable reason codes
Policy evidence is local only and does not grant payment authority

⚠️ Limitations

Not payment authorization: ELIGIBLE is policy evidence only
Does not include funds, private keys, payment credentials, or settlement authority
Requires a separate production payment adapter for payment-related capabilities

📌 Topics

ai-agents, mcp, model-context-protocol, policy-engine, typescript, spend-policy, deterministic, governance

Safety boundary

This package does not hold funds, private keys, payment credentials, customer data, or settlement authority. It does not sign, send, settle, custody, or record payments. It makes no network requests.

A production payment adapter needs separate, independently verified controls for authenticated policy/action provenance, canonical payload hashing, durable atomic budget reservation and replay protection, customer-controlled signing, rail validation, and settlement reconciliation.

Tool

evaluate_spend_policy accepts an evaluation time, a policy, a proposed action, and the already-spent amount. All money is passed as integer micro-unit strings, never JavaScript floating-point numbers.

The caller supplies the clock and already-spent value; therefore this tool is safe for dry runs and local evidence, not a replacement for a trusted payment or accounting system.

Agent contract

The versioned machine-readable contract is capabilities.json. It describes the only tool this package exposes, its required inputs, its three possible outcomes, and its non-negotiable safety boundary.

Transport: local stdio MCP

Tool: evaluate_spend_policy

Required inputs: evaluationTime, policy, action, and alreadySpentMicros

Outputs: ELIGIBLE, DENY, or STEP_UP, each with a stable reason code

Side effects: none

Credentials, payment execution, custody, signing, settlement, and network access: not supported

The manifest is package-source evidence for this release, not a promise of a hosted agent-discovery endpoint. ELIGIBLE remains local policy evidence only, not payment authorization.

io.github.dingdawg/agent-spend-policy-mcp