raposa-mcp

Raposa Aval is a human-approval MCP server by DC ESCRYPT (https://raposa.group). It gives an agent one honest capability: before a high-stakes action runs, a named human must explicitly approve it. The API is served from EU infrastructure; approvals are sealed in a hash-chained audit log that can be exported for an auditor.

https://dcescrypt.com · Documentation · Get a sandbox key

1 · Why fail-closed approval matters

request_human_approval(action, context, risk) returns:

{approved: true|false, decided_by, decided_at, comment, status, …}

approved is true only when an authorised human pressed Approve. Every other outcome — a rejection, an expiry, a timeout — returns approved: false. Silence is never consent:

The agent must branch on approved; the tool result never leaks a key or a secret. A decided approval is final: a second decision on the same id is refused with HTTP 409. Every decision lands in the hash-chained audit log (/v1/audit/verify, /v1/audit/export).

Related MCP server: gotoHuman MCP Server

2 · Architecture

Agent ──tool call──▶ MCP server (raposa-mcp) ──HTTPS, Bearer RAPOSA_API_KEY──▶ Raposa Aval API
                                                                             │
                                                          approval created (status: pending)
                                                                             │
                     human sees it: e-mail link · web console · Slack · Telegram
                                                                             │
     agent ◀── polls get_approval / webhook callback ◀── decision (approve|reject)

• The MCP server holds only the client API key. It reads RAPOSA_API_KEY from its environment and never includes it in tool inputs or outputs.

• Approvals are scoped per API key: one customer can never read or decide another customer's requests — foreign ids answer 404 Approval not found, and only an operator key can decide.

• Decisions carry the approver identity and a comment; re-deciding is refused.

3 · Getting an API key

1. Register a free sandbox at https://raposa.group/start/ (100 approvals a month).

2. An e-mail titled “Raposa — your sandbox key” arrives with a one-time claim link, valid 24 hours. The link shows the key once and cannot be reused; an expired or used link refuses with 410 gone.

3. The claim link is itself the credential: whoever opens it first within 24 h owns the key. Treat that mail as a secret — do not forward it or paste the link into tickets and chats. If the link may have leaked before you opened it, write to contact@raposa.group: an unclaimed link self-expires in 24 h and a claimed key can be revoked and reissued from the account page or by support.

4. Put the key into the MCP server’s environment (below).

No production data is needed for the sandbox; the same flow issues paid-plan keys automatically after checkout.

4 · Install (secure, step by step)

The server is a stdio MCP server, one command: uvx raposa-mcp (pin raposa-mcp==0.1.2 to freeze a published release).

4.1 The key never travels through a command line

Do not write RAPOSA_API_KEY=<value> into a chat, a shell command, a screenshot or a repository. Two supported ways to bring the secret to the MCP process:

Secret-manager export (1Password CLI, any similar tool works the same way):

# ~/.zshrc or ~/.bash_profile — resolved at login from your own vault
export RAPOSA_API_KEY="$(op read 'op://Dev-API/<your-item>/key')"

The expansion happens inside the shell that starts your MCP client; the key never appears as a literal argument (it would otherwise show in ps, in the shell history and in crash logs).

Plain-environment alternative: keep the value in a file with restrictive permissions and source it from your shell profile. Whatever the source, the constant rule is: the MCP process must inherit the variable from its parent environment, not receive it as an argument.

4.2 Claude Code

Claude Code supports ${VAR} expansion in MCP configuration — the placeholder is stored in the config file, resolved from the environment when the server starts (MCP docs):

claude mcp add raposa -e RAPOSA_API_KEY='${RAPOSA_API_KEY}' -- uvx raposa-mcp
claude mcp get raposa    # shows the ${...} reference — never the secret itself

Single quotes matter: without them the shell resolves the placeholder before Claude Code stores it. claude mcp get raposa is the check that no secret landed in the configuration.

4.3 Generic MCP stdio config (.mcp.json, other clients)

{
  "mcpServers": {
    "raposa": {
      "command": "uvx",
      "args": ["raposa-mcp"],
      "env": { "RAPOSA_API_KEY": "${RAPOSA_API_KEY}" }
    }
  }
}

4.4 Check it without spending the secret

# 1 · the stored placeholder is a reference, never the value:
claude mcp get raposa | grep -F '${RAPOSA_API_KEY}'     # prints the line — good
claude mcp get raposa | grep -F '<the key>'             # prints nothing — good
# 2 · end to end: list the tools (the server only runs when it can read the key):
#     three tools — create_approval, get_approval, request_human_approval
# 3 · a controlled read of a non-existent id must answer a 404 "Approval not
#     found", never a missing-environment error — that proves the key arrived

5 · Tools reference

All three tools take exactly these arguments (limits enforced client- and server-side):

create_approval

Create an approval request and return its id immediately (no waiting). Use with a webhook_url, or poll with get_approval.

Returns {"id": "…", "status": "pending"}.

get_approval

Returns the same shape for every terminal state:

{"id": "…", "status": "approved|rejected|expired|pending", "approved": true|false,
 "decided_by": "…", "decided_at": "…", "comment": "…", "message": "…"}

approved is true only when status == "approved".

request_human_approval

Same parameters as create_approval (no webhook_url), plus:

On a human decision returns the same shape as get_approval. On the deadline it returns {"status": "timeout", "approved": false, "id": "…", "message": "Not approved — nobody answered in Ns. The request is still open; check later with get_approval."} — and it really stays open until its TTL: decide late via your console or get_approval later.

6 · Worked examples

Real outputs (sandbox, uvx raposa-mcp==0.1.2):

create_approval → returns immediately:

{"id": "79d647c1-…", "status": "pending"}

get_approval while pending:

{"id": "79d647c1-…", "status": "pending", "approved": false,
 "decided_by": null, "comment": null, "message": "Still waiting for a human."}

a human approved (via e-mail link or console):

{"id": "79d647c1-…", "status": "approved", "approved": true,
 "decided_by": "<approver>", "decided_at": "2026-10-20T…", "comment": "owner mcp e2e approve",
 "message": "Approved by a human — proceed."}

a human rejected:

{"status": "rejected", "approved": false, "message": "Rejected by a human — do not proceed."}

TTL elapsed, nobody decided:

{"status": "expired", "approved": false,
 "message": "Not approved — the request expired before anyone decided."}

request_human_approval with wait_sec=8 and nobody answering:

{"id": "ed2e963e-…", "status": "timeout", "approved": false, "decided_by": null,
 "message": "Not approved — nobody answered in 8s. The request is still open; check later with get_approval."}

Error shapes you will actually see:

get_approval on a foreign or unknown id — existence is never confirmed to a foreign key:

Raposa API 404 on GET /v1/approvals/<id>: {"detail":"Approval not found"}

a decided approval does not flip: the second decision is refused by the API:

Raposa API 409 on POST /v1/approvals/<id>/decision

7 · Who holds which power

Nothing here makes an MCP client key an administrative credential. The administrative surface (operator console) is a separate, credentialed place — not reachable with an MCP key.

8 · MCP Rush hosted listing

The same server is published on MCP Rush hosted infrastructure: darius-ceponas/raposa-aval-human-approval-mcp — release 0.1.1, three tools, scanned and verified (trust score 96/100, grade A). The gateway is live and refuses calls that carry no buyer token.

Honest status: the hosted release answers, but an end-to-end gateway call with real buyer credentials has not been verified yet — the marketplace’s Access page (where a buyer mints the required gateway key) was unreachable at the time of writing, and the publisher has reported it to MCP Rush support. Until that lands, the direct uvx install in § 4 is the supported way; this README makes no gateway end-to-end claim.

9 · Troubleshooting

10 · Development

This repository — the MCP package itself:

pip install -e ".[test]" && pytest -q

11 contract tests run against a mocked transport: tools are listed with their schemas; approve / reject / timeout / expiry all read back fail-closed; create returns an id without waiting; API errors are reported, not swallowed; an unknown risk never reaches the API; a missing key is a clear error while the tool listing still works.

The backend API (quotas, key rotation, one-time claim links, tenant isolation, panel scopes, rate limiting, webhooks, Stripe, Slack/Telegram channels, audit chain) is a separate repository with its own suite; these pytest runs do not exercise it.

• Source and issues: https://github.com/agentlabbusiness/raposa-mcp

• Documentation: https://raposa.group/docs/

• Legal: MIT License — © DC ESCRYPT SL (see LICENSE)

• Contact: contact@raposa.group

Raposa Aval is a DC ESCRYPT product.