Self-hosted MCP gateway

A thin MCP server you run in your own network. Your AI assistant talks to it instead of scope.bid directly. The server forwards your requests to scope.bid and returns the results. Same dispatches. Same professionals. Same Stripe Connect payment rail. Your security team sees every call before it leaves your network.

Limited release. No general-availability date set.

Reference deployment

A reference deployment of the Scope gateway is live at scope-mcp-gateway.onrender.com. You can curl the health and tools endpoints from any environment to verify the protocol shape before deploying your own. This staging instance is rate-limited, sleeps on Render after 15 minutes of idle (the next call cold-starts), and rotates tokens nightly. Do not use it for production dispatches.

Check the reference deployment

curl https://scope-mcp-gateway.onrender.com/health
# -> JSON with "status": "ok", the vertical and the version

curl -H "Authorization: Bearer <token>" \
  https://scope-mcp-gateway.onrender.com/mcp/v1/tools
# -> JSON with the "tools" list

Architecture

Two outbound HTTPS connections leave your perimeter. No inbound traffic.

Your perimeter holds your AI assistant, the MCP tunnel you deploy, and the Scope gateway. Two outbound HTTPS connections leave it: the tunnel to your AI provider, the gateway to scope.bid. No inbound traffic.Your perimeterYour AI assistantMCP tunneldeployed by youScope gatewaystateless proxyAI providermodel, tunnelscope.bidprofessionalsHTTPSHTTPS
The tunnel and the gateway each open one outbound HTTPS connection. Nothing connects in.

Deploy

Provider-specific templates live in the scope-mcp repo under deploy/. Each one is a single-page README: prerequisites, deploy command, health-check verification.

Cloudflare Workers
deploy/cloudflare/README.md
Small (under 100 dispatches/day)
0.25 vCPU, 256 MB RAM
Medium (100-1,000 dispatches/day)
0.5 vCPU, 512 MB RAM
Large (1,000+ dispatches/day)
1 vCPU, 1 GB RAM

Auth and secrets

Scope issues a SCOPE_API_TOKEN at enterprise onboarding. The token is the gateway's only credential - injected at runtime via your provider's standard secret mechanism, never baked into the container image.

  • Rotation: restart the gateway after updating the secret.
  • Revocation: a revoked token is refused by scope.bid. The next upstream call fails 401.
  • Scope: tokens authorize the firm's organization on scope.bid. They never carry secrets that work outside that scope.

Observability

  • Logs: every tool call writes a structured JSON line to stdout - ts, level, msg, tool, upstream_status, latency_ms, and on successful calls token_suffix (last 4 only). Pipe stdout to your existing aggregator (Datadog, Splunk, Cloudflare Logs).
  • Metrics: Prometheus format at /metrics. Total tool calls, error count by upstream status, latency p50/p95/p99.
  • Health probes: /health for liveness (returns 200 + build version), /ready for readiness (returns 200 only after a successful upstream call).
  • Alerts to wire: upstream 5xx rate above 1%, p95 latency above 500ms, container restart loops.

Troubleshooting

  • Upstream 401: the SCOPE_API_TOKEN is expired or revoked. Generate a new token via Scope admin, inject via your secret mechanism, restart the gateway.
  • DNS resolution failures: the container needs egress to scope.bid over 443. Confirm with dig scope.bid inside the container.
  • TLS trust: the container ships with a stock CA bundle. If your network terminates TLS at a corporate proxy, mount your CA into /etc/ssl/certs and set NODE_EXTRA_CA_CERTS.
  • Rate limits: if the gateway returns 429, wait the Retry-After seconds and retry. For sustained 429s, email jack@scope.bid.

Status

Limited release. Currently available to qualifying enterprise customers. No general-availability date is set. Support during the limited release comes with no service-level commitment. For deployment scoping, token issuance, or production-readiness questions, email jack@scope.bid.