About SPR (Sealed Payment Record)
SPR is a local record system for sealing AI agent payment intent and settlement details with a company's key. It is for agent operators, finance teams and auditors who need independently verifiable payment evidence.
Key features
- 01Seal payment intent and settlement records
- 02Check payments against declared terms
- 03Export dispute evidence
- 04Verify sealed records with a public key
- 05Local books and x402 support
Use cases
- 01Record an agent's payment purpose and limit
- 02Check payee, amount and duplicate payments
- 03Prepare evidence for a chargeback
- 04Verify an exported payment book
MCP support
The local SPR MCP server exposes tools to record payment intent, check a payment against that intent, and record settlement. Records are sealed in a local book.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @luca-network/spr-mcpSPR_AGENT_IDThe agent's ID, such as agt_researchSPR_ENTITYThe company the agent pays forSPR_DIRThe book's folder (default ./spr-books); it holds the company's private keySPR_TIME_ZONEThe company's time zone for accounting dates (default UTC)SPR_BINDINGon, to return the nonce to sign with each x402 intentLocal servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
9Saved tool list · updated 1w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| rail | string | No | |
| product | string | No | Optional: what is being bought, such as an API plan, a dataset or a SKU |
| purpose | string | Yes | Why you are paying, in one sentence. This is what an auditor reads. |
| currency | string | No | |
| expires_in | string | No | ISO 8601 duration (PT1H, P3D) |
| counterparty | string | No | |
| max_amount_minor | integer | No | The most you may pay |
| expected_amount_minor | integer | No |
Details
From the project READMESPR: the Sealed Payment Record
Your AI agent pays for data, compute and services, and your finance team and auditor can't see what each payment was for. SPR records every payment your agent makes, before and after it pays, sealed with your company's own key on your own machine. Anyone can check the seals. No service has to be trusted, Luca's included.
npx -y @luca-network/spr-mcp
Free and open: Apache-2.0, with a royalty-free patent commitment (PATENTS.md).
Use it from your AI tool
Every setup below runs the same MCP server on your machine. Set SPR_AGENT_ID (your agent) and SPR_ENTITY (the company it pays for); records go to SPR_DIR.
Claude Desktop: add to claude_desktop_config.json:
{ "mcpServers": { "spr": { "command": "npx", "args": ["-y", "@luca-network/spr-mcp"],
"env": { "SPR_AGENT_ID": "agt_research", "SPR_ENTITY": "Example Co", "SPR_DIR": "/Users/you/spr-books" } } } }
Cursor: the same block in .cursor/mcp.json.
Claude Code:
claude mcp add spr -e SPR_AGENT_ID=agt_research -e SPR_ENTITY="Example Co" -- npx -y @luca-network/spr-mcp
OpenAI Agents SDK and LangChain (JavaScript):
import { SprBook, toOpenAIAgentsTools, toLangChainTools } from "@luca-network/spr-sdk";
const book = new SprBook({ dir: "./books", agentId: "agt_research", entity: "Example Co" });
const tools = toOpenAIAgentsTools(book, tool); // tool from "@openai/agents"; or toLangChainTools(book, tool)
CrewAI, Microsoft Agent Framework, Google ADK and the Claude Agent SDK (Python): each connects to the MCP server in a few lines. See examples/.
Any other MCP client: run npx -y @luca-network/spr-mcp over stdio with the two settings above.
The agent then calls record_intent before it pays, check_payment to see whether the payment is inside what it declared, goes to the payee it named and wasn't paid already, and record_settlement after it pays. Each record is sealed in the book's folder.
A merchant's book works the same way for payments received: it can keep the buyer's own sealed intent, how the buyer authenticated, and the order and processor references, and export one payment's evidence for a chargeback response (export_dispute_pack).
A record shows:
- what the agent declared before it paid: the purpose, the most it could spend, and when the permission expired;
- what was actually paid, on which rail, to whom and when;
- whether the payment stayed inside what was declared;
- for x402 payments, whether the chain itself recorded which declaration the payment was made under. The payer signs a fingerprint of the sealed intent as the payment's nonce.
This repository holds:
| Path | What it is |
|---|---|
SPEC.md | The format, draft 1 |
packages/spr | @luca-network/spr, the reference library for JavaScript and TypeScript, with no dependencies, plus the test vectors and spr-verify |
packages/spr-sdk | @luca-network/spr-sdk: seal your agents' payments locally, with withLuca(fetch) for x402 and tools for the OpenAI Agents SDK and LangChain |
packages/spr-mcp | @luca-network/spr-mcp: the same calls as an MCP server, for Claude Desktop, Cursor or your own agent (npx @luca-network/spr-mcp) |
python | spr-sdk for Python: the same format and the same books |
Seal an x402 payment in one line
import { SprBook } from "@luca-network/spr-sdk";
import { withLuca } from "@luca-network/spr-sdk/x402";
const book = new SprBook({ dir: "./books", agentId: "agt_research", entity: "Example Co" });
const pay = withLuca(book, { signer: account });
const res = await pay("https://data.example/prices", {
luca: { purpose: "Market data for the pricing model", max_amount_minor: 20000, currency: "USDC" },
});
res.luca.record; // sealed, matched to its intent, bound on chain
Check anything without Luca
npx spr-verify ./books-export --key company.pem # a package, and whose it is
python -m spr ./books-export --key company.pem
npx spr-verify --vectors # the library against the test vectors
python -m spr --vectors # the Python SDK against the same vectors
company.pem is the company's public key, obtained from the company and never from the package (book.publicKey() gives it). Without --key the checkers show only that a package is consistent with the keys it carries, and say so.
Security problems: see SECURITY.md.
An implementation in any language conforms when it gives the same answers as packages/spr/vectors/vectors.json.
Local SPR and hosted Luca
The packages here run on your own machine and make no call to Luca. Luca's hosted service is a separate product that a book can be sent to.
| Local SPR (these packages) | Hosted Luca | |
|---|---|---|
| Where it runs | Your machine, your folder, your key | Luca's service |
| What it records | The intent before paying and the payment after, as the agent reports them | The same, plus the payment as the rail itself reports it |
| Limits it checks | The declared intent: its cap, its expiry, whether it was withdrawn and the payee it named; the same invoice paid twice. Optionally the book's own mandate.json (limits, payees, a window for repeats), which whoever keeps the folder writes, so it is reported as authority: "self-declared" | Mandates your finance team approves in Luca |
| Confirmation from the rail | None: the record is what the agent reported | The payment read from the rail, and any difference flagged |
| Accounting | None | Classification to your accounts (US GAAP, IFRS, tax), journal entries posted to your ERP, the month's close |
| Evidence | Seals anyone can check with spr-verify or python -m spr | Evidence your auditor can check, and the month's audit package |
| Cost | Free, Apache-2.0 | A Luca account |
A book moves across with one call once finance has registered its key: book.sendToLuca(). Luca checks every seal before it stores anything.
Send the book to Luca (optional)
Finance registers the book's key with Luca once (book.registerWithLuca), and then book.sendToLuca() sends the book. Luca checks every seal before it stores anything, and sending the same book again changes nothing. Write to [email protected] for an account.
Changes
What changed in each version, and what you may need to change on upgrading: CHANGELOG.md.
Status
This is draft 1 of the format and version 0.1 of the libraries. We expect to change things as people use them. Tell us what breaks or what's missing: open an issue, or write to [email protected].
License
Apache License 2.0 for the code: see LICENSE. Luca's patent commitment, which lets anyone implement the format with or without this code, is in PATENTS.md.
Luca (luca.network) uses SPR for its hosted accounting of agent payments. That covers reading the rails, classifying to US GAAP and IFRS, posting to the ERP, labels agreed with counterparties, and the close.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.