Stipendex

Stipendex

Give your agent a stipend, not your credit card.

Open payment infrastructure for the agent economy: protocol + wallet + trust registry. Any LLM agent can safely pay any verified receiver — from half a cent for a paid MCP tool call to an $80 goods order.

Formerly working-titled ATP (Agent Transaction Protocol). The name is secured: stipendex.com · npm stipendex


The idea in three sentences

Agents have learned to act (MCP gave them hands) but not to pay: there is no standard way to delegate money to an agent with verifiable boundaries. Stipendex defines a minimal payment cycle (Quote → Proof → Receipt), a deterministic wallet kernel with a signed mandate — the only place where a payment decision is made (the LLM never touches money) — and a trust registry that answers "who may be paid at all". Settlement is executed by existing rails (Stripe, x402, ledger accounts) through adapters — Stipendex builds no new rails and competes with none.

Components

Component What it is Openness
Stipendex specification Quote/Proof/Receipt envelope, mandate, transports (MCP, HTTP), profiles open spec
@stipendex/wallet wallet kernel: SDK for agent hosts (limits, HITL, keys) open source
@stipendex/server paywall middleware for MCP servers and HTTP APIs open source
Trust registry signed descriptors of verified receivers; federated providers open format; providers are commercial
Facilitator accounts, payment verification, HITL queue, anti-fraud open reference API; packages/facilitator is archived (see the on-chain profile)
On-chain profile LMB token (EIP-3009), faucet with gas stipend, Market contract, catalog and relay at 4agnts.com open source, live on Base Sepolia

Three profiles

Relationship to 4agents

4agents (a sibling project) is a trust layer for agentic shopping. The projects complement each other and share components:

Details: docs/product.md.

Document map

Document Contents
docs/product.md problem, positioning, 2026 landscape, distribution strategy, monetization
docs/spec/overview.md roles, objects, invariants, adoption levels L0–L3
docs/spec/envelope.md Quote / Proof / Receipt, tolerance, tab sessions, idempotency
docs/spec/mandate-wallet.md mandate, wallet kernel, HITL modes, Confirmation
docs/spec/transports.md bindings: MCP (-32402) and HTTP (402)
docs/spec/rails.md rail adapters: account, stripe-mpp, stripe-connect, x402
docs/spec/checkout-profile.md goods profile: hold→ACCEPT→capture, 4agents integration
docs/spec/onchain-profile.md on-chain profile (live): LMB, faucet, Market, delivery envelopes, relay and catalog APIs
docs/security.md threat model and normative requirements
docs/trust-registry.md trust registry: descriptors, API, federation
docs/roadmap.md phases, publication, licenses, MCP SEP
docs/ru/ Russian mirror of this documentation
CONCEPT.md archive (Russian): original concept v1 → critique → v2 (decision history)

Principles (short version)

  1. The LLM never touches money. Payment decisions are made by a deterministic wallet against a signed mandate; the model has no "transfer to X" operation.
  2. A thin layer, not a sixth protocol. The envelope and the authority model are ours; the rails are others' (Stripe, x402, ledger accounts); commerce scenarios are others' (ACP/UCP/4agents).
  3. Works out of the box. The default account rail plus a hosted facilitator: a server monetizes in ~10 lines, an agent connects a wallet in ~5.
  4. Aggregation by default. Micropayments accumulate on a session tab and settle on a threshold or session close; per-call settlement is an option, not the default.
  5. Default deny. Payments go only to receivers on the allowlist / trust registry; a new receiver always requires human confirmation.
  6. Honest amounts. When the final amount is unknown — hold on estimate+tolerance, capture only the confirmed amount (≤ hold). No blank checks.