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.

Related MCP server: hivelaw

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

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

Related projects

License

Apache-2.0