CAIN-42 CAIN Studio
CAIN-42

Deployment + adversarial verification · 10 business days

MCPGate Assessment & Remediation

We deploy the enforcement boundary in a staging environment you control, build a default-deny policy from your real tool surface, and prove with a 19-case harness that denied calls never reach the tool.

Request this engagement — $2,500 50% deposit. Scope confirmed after a 30-minute technical call.

Who this is for

A team that wants the gate in its own environment — not a hosted demo — or already runs a proxy it cannot prove stops anything. You have a staging cluster, Docker host, or a namespace you can deploy into.

What we do

StepOutput
Baseline the current pathWhat reaches the tool today, what is logged, what is only assumed.
Deploy deploy/mcpgate-standaloneUnprivileged container, read-only root, default-deny config, persistent nonce store, tool pinning.
Build the policyPer-tool allow / approval / deny, derived from your tool surface, reviewed with you.
Run the harnessAllowed, denied, unmapped, approval, malformed, replay, expiry, wrong-tool, wrong-arguments, direct bypass, downstream failure, evidence tamper.
Hand overSigned bundle, verification report, runbook, backup/restore, upgrade/rollback, uninstall, acceptance checklist.

What you get

Deployed gate + config

A running boundary in your staging environment, with the policy report and the exact image digest.

Verification report

verification-report.json: every case, expected vs observed, and the downstream call log proving denied tools were not reached.

Evidence bundle

Signed, chained, offline-verifiable. Includes the tamper case: a modified bundle must be rejected.

Operations pack

Runbook, backup/restore, upgrade with rollback, uninstall, handoff with limitations and sign-off.

Disclosed limitations

  • Network bypass is real. The gate governs only traffic routed through it. If the agent can reach the tool server directly, it is outside the boundary. Closing that route is part of the engagement, and it is your network policy that enforces it.
  • One gate, one downstream, one host in the standard package. Multi-replica single use requires a shared nonce database; scaling out is a separate scope.
  • No third-party audit. The 19-case harness is reproducible evidence, not an independent assessment.
  • Not production operations. We hand over a staging deployment and the procedures. Running it 24/7 is yours or a separate retainer.

What we need from you

  • A staging environment and permission to deploy a container.
  • The real tool definitions and a test agent that can call them.
  • A named owner for the network boundary and the tool server.
  • Written authorization for the work.

Request it

We confirm the environment and the tool surface before quoting. If the scope does not fit this fixed price, we will say so rather than deliver something smaller and call it complete.

Request the $2,500 engagement

Prefer email? vincent@firmeai.com