F
filippoplesca3-wq/MandateAgent
Signed spending authority for AI agents on PayPal
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
官方说明Signed spending authority for AI agents on PayPal. Ed25519 mandates, a deterministic guard, and an append-only audit ledger. PayPal AI Hackathon 2026.
安装与配置
取自该仓库的安装命令
npm install cp
解析自我们于 2026-11-09 抓取的该仓库 README;我们未执行过。
这条是仓库自己写的安装方式,不是 MCP 客户端配置,所以不生成 JSON。
配置项的键名以你的客户端为准:客户端与 CLI
- ★ 数取值
- 0 ★ · filippoplesca3-wq/MandateAgent · 取自 GitHub 搜索 API · (2026-11-13 20:00:50 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-10 · 最近见到 2026-11-14 · 共见到 5 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 个来源 · repo
- 安装
npm install cp
常见问题(答案都来自我们的记录)
怎么安装?
该仓库给出的安装命令是
npm install cp。我们的记录里它是仓库自己提供的;不保证在你的客户端上直接可用。它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
filippoplesca3-wq/MandateAgent;我们收录于 2026-11-10;最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。# Agent Mandate
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
Signed spending authority for AI agents on PayPal.
Agent Mandate lets an AI agent buy things for a person without ever holding the
wallet. The human signs a scoped, expiring, revocable mandate — total budget,
per-payment ceiling, merchant allowlist and an approval threshold — and every
PayPal tool call the model makes passes through a deterministic guard that
resolves the real order amount and merchant, checks it against the signed claims,
records the decision in an append-only ledger, and refuses anything outside the
bounds. The model can be prompt-injected, confused or simply wrong; the mandate
still holds. The hosted demo replays a real Gemini run, and a judge can press
Run live with Gemini to watch the real model run now against simulated
PayPal. The local live path creates and captures real PayPal sandbox orders.
The console after one recorded Gemini run: $24 allowed, $45 escalated and
approved by a human, $99 denied, and every decision in the ledger.
| Live demo | https://agent-mandate.onrender.com (replace after deploy) |
| Video (under 3 min) | https://youtu.be/... (placeholder — public YouTube link) |
| Source | https://github.com/filippoplesca3-wq/MandateAgent |
| License | MIT |
## Judge in 90 seconds
1. Open the hosted demo and press Play recording. The header says
replay; the banner names the model and the capture date. No key, no API
calls. If the instance is asleep, npm ci && npm run demo serves the same
recording at http://localhost:3001.
2. Or run the model live. Press Run live with Gemini in the banner. The
server runs a real Gemini agent against simulated PayPal in a throwaway
store, then shows the fresh mandate, the guarded tool calls, the $99 denial
and a server-computed signature check. No PayPal credentials are involved
and no money can move. Runs are capped per day and per visitor.
3. Read the ledger. One run produces all three outcomes: $24 allowed, $45
escalated and approved, $99 denied at sketchy-deals.
4. Try a policy check. The Test a payment presets run a real dry-run
evaluation against the pre-seeded mandate.
5. Verify the mandate. Press Verify next to the signature. The console
shows SIGNATURE VALID · Ed25519 · recorded key: the mandate on screen
checks out against the key the recording was signed with.
6. Run the proof. npm ci && npm test (206 tests),
npm run verify:recording (offline Ed25519 check, no server),
npm run check:demo -- (audits the demo's own claims) and
npm run check:live (real PayPal sandbox + Gemini).
7. Go deeper: architecture,
security model, Devpost copy,
demo script.
If the hosted instance is asleep, the first request can take 30–60 seconds. The
Render blueprint deploys a fresh copy with no credentials, and
npm run demo runs the recording locally with no credentials at all.
## How PayPal is used
Every spend goes through the real PayPal Agent Toolkit (47 tools exposed by
@paypal/agent-toolkit), not a hand-written HTTP wrapper.
The live path creates and captures real sandbox orders: create_order →
payer approval in the browser → pay_order → get_order.
The browser checkout uses the PayPal JS SDK v6 with the sandbox client id;
only the public client id reaches the browser.
Agent Mandate lets an AI agent buy things for a person without ever holding the
wallet. The human signs a scoped, expiring, revocable mandate — total budget,
per-payment ceiling, merchant allowlist and an approval threshold — and every
PayPal tool call the model makes passes through a deterministic guard that
resolves the real order amount and merchant, checks it against the signed claims,
records the decision in an append-only ledger, and refuses anything outside the
bounds. The model can be prompt-injected, confused or simply wrong; the mandate
still holds. The hosted demo replays a real Gemini run, and a judge can press
Run live with Gemini to watch the real model run now against simulated
PayPal. The local live path creates and captures real PayPal sandbox orders.
The console after one recorded Gemini run: $24 allowed, $45 escalated and
approved by a human, $99 denied, and every decision in the ledger.
| Live demo | https://agent-mandate.onrender.com (replace after deploy) |
| Video (under 3 min) | https://youtu.be/... (placeholder — public YouTube link) |
| Source | https://github.com/filippoplesca3-wq/MandateAgent |
| License | MIT |
## Judge in 90 seconds
1. Open the hosted demo and press Play recording. The header says
replay; the banner names the model and the capture date. No key, no API
calls. If the instance is asleep, npm ci && npm run demo serves the same
recording at http://localhost:3001.
2. Or run the model live. Press Run live with Gemini in the banner. The
server runs a real Gemini agent against simulated PayPal in a throwaway
store, then shows the fresh mandate, the guarded tool calls, the $99 denial
and a server-computed signature check. No PayPal credentials are involved
and no money can move. Runs are capped per day and per visitor.
3. Read the ledger. One run produces all three outcomes: $24 allowed, $45
escalated and approved, $99 denied at sketchy-deals.
4. Try a policy check. The Test a payment presets run a real dry-run
evaluation against the pre-seeded mandate.
5. Verify the mandate. Press Verify next to the signature. The console
shows SIGNATURE VALID · Ed25519 · recorded key: the mandate on screen
checks out against the key the recording was signed with.
6. Run the proof. npm ci && npm test (206 tests),
npm run verify:recording (offline Ed25519 check, no server),
npm run check:demo -- (audits the demo's own claims) and
npm run check:live (real PayPal sandbox + Gemini).
7. Go deeper: architecture,
security model, Devpost copy,
demo script.
If the hosted instance is asleep, the first request can take 30–60 seconds. The
Render blueprint deploys a fresh copy with no credentials, and
npm run demo runs the recording locally with no credentials at all.
## How PayPal is used
Every spend goes through the real PayPal Agent Toolkit (47 tools exposed by
@paypal/agent-toolkit), not a hand-written HTTP wrapper.
The live path creates and captures real sandbox orders: create_order →
payer approval in the browser → pay_order → get_order.
The browser checkout uses the PayPal JS SDK v6 with the sandbox client id;
only the public client id reaches the browser.
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/filippoplesca3-wq-mandateagent/)
- 来源
- github · https://github.com/filippoplesca3-wq/MandateAgent
- 按领域浏览
- 开源项目
- 收录于
- 今天
同类推荐
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。