@humanaccepted/mcp

@humanaccepted/mcp

MCP server that auto-emits tamper-evident receipts for every tool call, enabling EU AI Act Article 12 compliance with signed, chain-linked receipts.

Category
访问服务器

README

<!-- Brand: HumanAccepted · color #0E5C4A · "Signed evidence your AI did what you said." -->

humanaccepted-mcp

MCP server that emits signed, EU AI Act Article 12-compliant receipts on every AI tool call.

Drop-in Model Context Protocol server. Every tools/call your agent makes — through Claude Desktop, Cursor, or any MCP-compatible host — is wrapped and emitted as a signed tool_call receipt into the HumanAccepted registry, chained to the previous receipt in the session, and verifiable offline against the public Python / TypeScript reference verifiers.

A single line of MCP config, no code changes, no host rewrites.


What it does

  1. Sits in front of your agent's tool layer as a passthrough MCP server.
  2. Wraps every tools/call in a before → after pair of signed receipts:
    • Start receipt (kind: "tool_call", duration_ms: 0) — captures intent + inputs.
    • End receipt — captures output + duration + ok | error | timeout | denied result.
  3. Chains both via chain.prev so an auditor can replay the full session in order.
  4. Generates a stable run_id per MCP session for end-to-end traceability.

All receipts are signed with the tenant's Ed25519 key inside the HumanAccepted Worker. The MCP server itself holds no private key material — it authenticates with Bearer sk_... and the Worker does the signing.


Quick start

# 1. Get an API key (one-time, returns a tenant + sk_…)
curl -X POST https://humanaccepted.ola-turmo.workers.dev/v1/bootstrap

# 2. Run the MCP server (npx pulls it from npm; no local install needed)
HUMANACCEPTED_API_KEY=sk_... \
  npx -y @humanaccepted/mcp

To use it in Claude Desktop or Cursor, add this to your MCP config:

{
  "mcpServers": {
    "humanaccepted": {
      "command": "npx",
      "args": ["-y", "@humanaccepted/mcp"],
      "env": {
        "HUMANACCEPTED_API_KEY": "sk_...",
        "HUMANACCEPTED_AGENT_ID": "my_agent",
        "HUMANACCEPTED_AGENT_TYPE": "claude-sonnet-4.6"
      }
    }
  }
}

The server connects on stdio, prints [humanaccepted-mcp] Connected to … to stderr, and from that point every tool call your agent makes is recorded.


Configuration

Env var Required Default Purpose
HUMANACCEPTED_API_KEY yes Tenant sk_… from POST /v1/bootstrap
HUMANACCEPTED_BASE_URL no https://humanaccepted.ola-turmo.workers.dev Override for staging / local Worker
HUMANACCEPTED_AGENT_ID no mcp_agent Stable id across the session (good for dashboards)
HUMANACCEPTED_AGENT_TYPE no mcp e.g. claude-sonnet-4.6, openai:gpt-5.5, cursor-2.0

What it exposes

This MCP server exposes one meta-tool:

  • humanaccepted_session_info — returns the current session's run_id, last receipt id, base URL, agent id, and tenant key fingerprint. Useful for debugging and for letting downstream code correlate receipts to a session.

All other tools/call requests are passed through but wrapped: the upstream tool still runs as normal, and a tool_call receipt pair is emitted to HumanAccepted for each invocation.


Receipt shape

Receipts follow HumanAccepted spec v1.1 (kind: "tool_call" was added in v1.1). A minimal example of the before half:

{
  "kind": "tool_call",
  "version": 2,
  "tenant": "ten_3K9F2X",
  "id": "rcp_4PZ8QJ2M1X9Y7K0ABCDEF1234",
  "issued_at": "2026-06-28T09:53:11.421Z",
  "agent": {
    "id": "my_agent",
    "type": "claude-sonnet-4.6",
    "run_id": "run_01HF3R8K2Q5P9X7Z4M6V0YJWBN"
  },
  "tool": {
    "name": "get_weather",
    "version": "1.0.0",
    "input": { "city": "Oslo" },
    "output": null,
    "duration_ms": 0,
    "result": "ok"
  },
  "chain": { "prev": null, "run_id": "run_01HF3R8K2Q5P9X7Z4M6V0YJWBN" },
  "canonical": "ha:v1:tool_call:ten_3K9F2X:rcp_4PZ…:{…sorted-utf8-json…}",
  "signature": {
    "alg": "ed25519",
    "key_id": "k_8N4Q",
    "value": "7c3f…(ed25519 sig over canonical)"
  }
}

The after half repeats the same envelope with duration_ms > 0, the actual output, and chain.prev = rcp_4PZ… (the start receipt's id).

Verify offline:

# Python
python3 -m verifier.python.verify --receipt rcp_4PZ... --pubkey <tenant-ed25519-pub>

# TypeScript
npx @humanaccepted/verify --receipt rcp_4PZ... --pubkey <tenant-ed25519-pub>

EU AI Act Article 12

Article 12 of the EU AI Act requires providers of high-risk AI systems to keep automatic, tamper-evident logs of every interaction over the system lifecycle. The deadline is 2 August 2026 — 35 days from today (2026-06-28).

HumanAccepted is the open primitive for Article 12. The MCP server here is the drop-in layer: any MCP-compatible agent — Claude Desktop, Cursor, Claude Code, your custom host — gets Article 12 logging with one block of MCP config and zero changes to the agent code.

After running your agent for any length of time, fetch the full session's receipts:

curl https://humanaccepted.ola-turmo.workers.dev/v1/article12-export \
  -H "Authorization: Bearer $HUMANACCEPTED_API_KEY" \
  | jq '.counts, (.receipts | length)'

Or replay one session:

curl https://humanaccepted.ola-turmo.workers.dev/v1/runs/$RUN_ID \
  -H "Authorization: Bearer $HUMANACCEPTED_API_KEY" \
  | jq '.chain_valid, .count'

Local development

The repo ships TypeScript source in src/ (entry: src/server.ts) and a built dist/ for npx.

git clone https://github.com/Ola-Turmo/humanaccepted-mcp
cd humanaccepted-mcp

# Workaround for openclaw-user npm issues on shared VPS:
npm install --prefix /tmp/mcp-dev --cache /tmp/mcp-dev/.npm-cache
NODE_PATH=/tmp/mcp-dev/node_modules node --import tsx src/server.ts

Or build:

npm run build && npm start

Tests:

HUMANACCEPTED_API_KEY=sk_... \
  HUMANACCEPTED_BASE_URL=http://127.0.0.1:8789 \
  npm test

Repo layout

Path Purpose
src/server.ts MCP server implementation — wraps tools/call, emits receipts
src/cli.ts Thin entrypoint (#!/usr/bin/env node)
dist/ Built JS for npx + bin: humanaccepted-mcp
test/server.test.mjs Boot + tools/list smoke test
package.json @humanaccepted/mcp v0.1.0, Apache-2.0
tsconfig.json ES2022, strict, outDir dist/

Related projects

Project What it is
Ola-Turmo/humanaccepted-spec Open spec v1.1 — receipt shapes, canonical form, verifiers in 5 languages
Ola-Turmo/humanaccepted Host product — Worker + Pages app (PR #6 ships v0.3 with these receipt kinds)
Live Worker https://humanaccepted.ola-turmo.workers.dev/ — the registry this server writes to
Live host product https://humanaccepted-new.pages.dev/ — dashboard + docs UI
Landing page (this repo) site/index.html — single-file CF Pages marketing page

License

Apache-2.0

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选