
Agent Receipts Mcp
io.github.dhanushs1912-svg由 GitHub 账号 dhanushs1912-svg 发布v0.1.2更新于 Oct 24, 2026
EU AI Act-ready audit trail: signed, tamper-evident receipts proving what your AI agents did.
概览
一个本地 MCP 服务器,为 AI 智能体的操作签发加密签名、哈希链式回执,并导出带签名的审计包供合规审查。
- 可以做什么
- 签发带签名和时间戳的回执,记录操作类型、摘要、执行者与主体,载荷仅保留 SHA-256 指纹后即被丢弃,不予存储。校验其他智能体或系统提交的回执,依据内嵌的 Ed25519 签名判断是否被篡改。按日期范围导出带签名的审计包,内含标记序号缺口与链断裂的防篡改报告。获取服务信息,例如用于离线校验的公钥。
- 适用场景
- 适用于智能体执行需要日后举证的重要操作(如审批、付款、记录变更),且需要向审计人员、监管机构或交易对手提供防篡改证据的场景,包括欧盟 AI 法案的记录留存讨论。也适用于核实其他智能体提交的回执是否被改动。
- 运行要求
- 通过 npx 以 stdio 在本地运行,需要 Node.js 22.13 或更高版本(使用内置 node:sqlite)。环境变量 AGENT_RECEIPTS_DATA 指定存放 SQLite 保险库与签名密钥的目录,默认 ./data。无需账号或网络访问,校验可离线完成。
安装
在 SourceWeft 中
- 打开 控制台中的 Agent Receipts Mcp,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
agent-receipts
Prove what your AI agents did.
On August 2, 2026, the EU AI Act's record-keeping obligations for high-risk AI systems became enforceable: automatic event logging (Article 12), deployer log retention of at least six months (Article 26), and penalties up to €15M or 3% of global turnover. The question your auditor, regulator, or counterparty will ask is simple:
"Your AI agent approved this. Prove exactly what it did — and prove nobody edited the record."
An ordinary log file can't answer that: whoever owns the log can edit it. agent-receipts
gives every consequential agent action a cryptographically signed, timestamped receipt
that anyone can verify offline — and exports signed audit bundles you can hand
directly to a compliance officer.
What a receipt proves
- What happened, when — action type, summary, actor, principal, ISO 8601 timestamp
- It hasn't been altered — Ed25519 signature over canonical JSON; change one character and verification fails
- Nothing was deleted — receipts are hash-chained with monotonic sequence numbers per vault; every audit export automatically flags gaps (deletions) and chain breaks
- Without exposing your data — hash-only privacy: payloads are SHA-256 fingerprinted and immediately discarded, never stored or transmitted
Built for the audit conversation
export_audit_bundle produces what compliance actually needs: all receipts in a date
range, a tamper-evidence report (sequence gaps, chain breaks), the signing public key,
and a manifest signature over the whole bundle. Retention is yours to control — the vault
is a local SQLite file you keep as long as Article 26 (or your policy) requires.
You don't have to trust us. Receipts verify offline against a pinned public key with the bundled CLI — no account, no network call, no dependence on this service existing tomorrow. That's what makes the evidence durable.
Quick start (MCP — local, free)
Requires Node.js >= 22.13 (uses the built-in node:sqlite — zero native dependencies).
Prefer one-click? Grab the .mcpb bundle from
Releases and open it
with Claude Desktop.
Quick start (HTTP API)
Service manifest for machine discovery: GET /.well-known/agent-receipts.json.
Offline verification
Receipt format (agent-receipts/v1)
Canonicalization is JCS-style: lexicographically sorted keys, no whitespace. The JSON Schema lives at docs/receipt.schema.json.
EU AI Act mapping (informational)
Not legal advice. Whether your system is "high-risk" and what you must log is a question for your counsel. agent-receipts produces evidence infrastructure.
Status & roadmap
- Core: issue / verify / export with hash chain + Ed25519
- MCP stdio server + HTTP API + offline verifier CLI + MCPB bundle
- Hosted vaults with per-receipt pricing
- RFC 3161 / transparency-log timestamp anchoring
- Key rotation and multi-key verification
MIT License.
来源:README.md,提交 3e40a58
工具
0版本历史
1- v0.1.2最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分