invinoveritas — MCP server

An independent review gate for AI agents, as a hosted MCP server. Before your agent takes an irreversible action — shipping code, running a shell command, signing a transaction — it calls review for a neutral second opinion from a party that isn't the agent, and gets back a recomputable, Bitcoin-anchored proof anyone can re-verify without trusting the agent or us.

The one thing an agent can't self-serve is a trustworthy verdict on its own output. This is that verdict — model-agnostic, advisory (it never blocks), and checkable from the bytes.

• Hosted, no install. Remote Streamable-HTTP MCP server: https://api.babyblueviper.com/mcp/verify
• Sign-in: standard MCP OAuth in clients that support it (a free account can be created on the consent screen), or a free API key as a Bearer header: curl -X POST https://api.babyblueviper.com/register -d '{"label":"mcp"}'
• verify_proof and ledger are always free / no auth.

Tools

Tool	What it does	Auth
review	Independent verdict (approve / approve_with_concerns / reject) on a code diff, shell command, plan, config, trade, or on-chain transaction, with ranked issues + a portable signed proof.	Bearer (paid)
verify_proof	Recompute and verify a signed invinoveritas proof someone handed you — confirms it without trusting the presenter or us.	free
witness	Anchor a third party's exact claim bytes, unmodified and unjudged — pure notarization (receipt + timestamp + integrity), distinct from review (our verdict). Source is self-declared and marked as such in the signed proof.	Bearer (paid)

Also on this endpoint: ledger (free public verdict track record), validate (backtest reality-check), audit_agent_readiness, ledger_submit, conformance_certify. 8 tools in all; the server never moves money or assets for the user.

Add it to your agent

Cursor — one-click: Add invinoveritas to Cursor (it signs in with OAuth on the first paid call), or edit ~/.cursor/mcp.json (global) / .cursor/mcp.json (project):

{
  "mcpServers": {
    "invinoveritas": {
      "url": "https://api.babyblueviper.com/mcp/verify",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

Cline — Add via the MCP marketplace, or in cline_mcp_settings.json:

{
  "mcpServers": {
    "invinoveritas": {
      "type": "streamableHttp",
      "url": "https://api.babyblueviper.com/mcp/verify",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

Gemini CLI — install as an extension:

gemini extensions install https://github.com/babyblueviper1/invinoveritas-mcp

Claude Code:

claude mcp add --transport http invinoveritas https://api.babyblueviper.com/mcp/verify \
  --header "Authorization: Bearer YOUR_API_KEY"

Devin — Settings → Connections → MCP servers → Add a custom MCP → HTTP (Streamable), URL https://api.babyblueviper.com/mcp/verify, Auth Header Authorization: Bearer YOUR_API_KEY.

OpenHands — config.toml:

[mcp]
shttp_servers = [{ url = "https://api.babyblueviper.com/mcp/verify", api_key = "YOUR_API_KEY", timeout = 60 }]

Any MCP client — point it at https://api.babyblueviper.com/mcp/verify (Streamable HTTP) with header Authorization: Bearer YOUR_API_KEY. Protocol 2025-06-18; the handshake, tools/list, and tools/call work out of the box.

Use it

Tell your agent: before opening a PR or running an irreversible command, call review with the diff/command as artifact, set artifact_type (code_diff / shell_command / onchain_action / general), and sign: true. On reject, hold and fix; on approve, attach the proof so reviewers can re-verify it.

Pairing with a fast gate (e.g. Jev)

Fast judgment models such as TypeSafe's Jev (jev_gate in the community jev-mcp server) screen every tool call in a few hundred milliseconds and answer allow / confirm / block. They are a good first line. What they don't give you is an independent verdict someone else can check later. Use both: let the fast gate handle the routine calls, and escalate only the risky or irreversible ones to invinoveritas.

gate = jev_gate(action=cmd, context=why)             # ~0.3 s, every call
if gate.recommendation == "allow" and not gate.touches_prod:
    run(cmd)                                          # routine: no second opinion needed
elif gate.recommendation == "block":
    stop_and_explain(gate)                            # clear no
else:                                                 # "confirm", or anything touching prod
    v = review(artifact=cmd, artifact_type="shell_command", context=why, sign=True)
    if v.verdict == "reject": stop_and_explain(v.issues)
    else: run(cmd); keep(v.proof)                     # signed proof anyone can verify_proof later

The fast gate keeps latency and cost low on the many calls; the signed verdict covers the few that matter, and leaves a record that doesn't depend on trusting either the agent or us.

Why a verdict you can recompute

review with sign: true returns a proof that is a BIP-340-signed Nostr event, published to a public, Bitcoin-anchored ledger. Verify it offline against our published key — recompute the event id and check the schnorr signature — or POST it to https://api.babyblueviper.com/verify-proof (free, no auth). Trust the math, not us. Wins and losses are published; the track record is a forensic artifact, not a score.

• Recompute a proof yourself: zero-dep pip install invinoveritas-verify / npm i invinoveritas-verify
• Server card: https://api.babyblueviper.com/.well-known/mcp/server-card.json

License

MIT