Proctor MCP server

The human-in-the-loop that MCP's specification asks for and does not provide.

The MCP tools spec says, verbatim:

"For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations."

Applications SHOULD ... "Present confirmation prompts to the user for operations, to ensure a human is in the loop."

It then defines no mechanism for it. Every client builds its own confirmation dialog, locally, and nothing survives the click. Three questions an auditor asks after an incident have no answer:

| Question | What a local dialog can show | |---|---| | Was a human actually there? | Nothing. A dialog proves a click, not a person | | Was it someone other than the agent's operator? | Nothing | | Can you prove it to someone who does not trust you? | Nothing |

Meanwhile the direction of travel for paid MCP tools is explicitly the other way. Vercel's x402-mcp pitch is "no account, no API key, no human in the loop". That is correct for a weather API and wrong for a wire transfer.

This server is the missing mechanism.

What it does

One tool: request_human_approval.

It blocks the agent on a real person who is not the operator, fails closed on a deadline, and returns a citation on a Hedera topic with no admin key — so the decision can be proven to a third party later, without trusting the agent or us.

below threshold   → Proceed. Policy did not require a human (below_threshold).
above threshold   → blocks, hands a link to a phone, waits
approved          → Approved by a verified human. HCS sequence #41. You may proceed.
refused / expired → NOT approved (EXPIRE). Do not proceed.

Most calls never reach a human. An oversight tool that interrupts everything is one the agent's owner switches off, so policy runs first and the common path is free.

Publishing status

Not published yet. The package is built, node-tested and manifest-valid, but the three remaining steps all need accounts only you can log into.

# 1. npm (needs `npm adduser` first)
cd mcp && npm run build && npm publish --access public

# 2. mcp-publisher
brew install mcp-publisher          # or the release tarball
mcp-publisher login github          # device code, needs the GitHub account
mcp-publisher publish               # reads server.json

# 3. verify
curl "https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.louissarvin/proctor"

The registry hosts metadata only, so npm must come first. Names are already aligned: mcpName in package.json equals name in server.json equals io.github.louissarvin/proctor, which the registry requires and which GitHub auth requires to match your username.

One real blocker beyond accounts: repository.url points at github.com/louissarvin/proctor, and that repo is not pushed. Publish after the push, not before.

Install

// Claude Desktop: claude_desktop_config.json
{
  "mcpServers": {
    "proctor": {
      "command": "bun",
      "args": ["/absolute/path/to/proctor/mcp/src/index.ts"],
      "env": { "PROCTOR_API": "http://localhost:3700" }
    }
  }
}

Works with any MCP client — Claude Desktop, Cursor, or your own. Transport is JSON-RPC 2.0 over stdio, protocol version 2025-06-18.

Verify it without a client:

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | bun src/index.ts

Three decisions worth knowing

A refusal is not a tool error. It returns isError: false with approved: false. Marking a refusal as an error invites a model to treat it as a transient fault and retry around it, which is precisely the behaviour this tool exists to prevent.

Unreachable is not refused. If Proctor is down the tool says so explicitly and marks isError: true. "The gate is down" and "a human said no" are different facts, and collapsing them means an outage silently becomes a rejection — or worse, gets retried until it isn't.

No dependencies on the oversight path. JSON-RPC over stdio is a loop, and adding an SDK to a security tool is how it acquires a supply chain. The x402 client is imported lazily and only when a wallet is configured, so pointing this at a hosted Proctor to ask a human a question never loads a Hedera SDK.

Paid mode

Set a wallet and the tool stops simulating and starts paying:

"env": {
  "PROCTOR_API": "http://localhost:3700",
  "HEDERA_AGENT_ID": "0.0.xxxxx",
  "HEDERA_AGENT_KEY": "302e…",
  "AGENT_ALLOWED_TOKENS": "0.0.10394781"
}

It then hits the real x402 gate and settles through the Blocky402 facilitator. A verified run:

[proctor] paid the gate; a witness is being dispatched from the rota
structuredContent: { "approved": false, "outcome": "EXPIRE", "paid": true }

and the corresponding settlement on Hedera testnet:

0.0.10359475   -420000     the agent pays
0.0.10349677  +415800     treasury
0.0.10349667    +4200     1% HTS custom fee, assessed on chain
fee payer 0.0.7162784     Blocky402

Policy runs first, and it is free. POST /v1/gate/evaluate decides whether a human is needed before any payment is made. An oversight tool that charges on every invocation is one the agent's owner switches off within a day.

Paid mode returns no witness link, on purpose. The gate never hands the approval URL back to the payer — an agent that could reach it could approve its own decision, which is the single property this product sells. The human is reached through the operator's rota instead. Unpaid operator mode does return a link, because there the caller is the operator.

Without a wallet the tool falls back to Proctor's operator endpoint, which opens a real decision but settles nothing, and says so in its own output (paid: false).

What it does not do yet

The evidence citation may read attestation pending when a decision resolves faster than consensus. The record lands afterwards, off the agent's critical path, and the sequence number is durable in the decision's own record.

Only the Hedera rail is wired here. The gate advertises Arc through Circle's Gateway too, and agent/ demonstrates it; the MCP client registers the Hedera scheme only.