Skip to content

Meet DeepintShield at Bengaluru Tech Summit 2026, 17 to 19 November, BIEC Hall 1, booth H1A106. Book 20 minutes

Start For Free

Platform · MCP Gateway

MCP gateway: every MCP server behind one governed boundary

An MCP gateway sits between AI agents and Model Context Protocol servers and decides which tools each caller can see and run. DeepintShield acts as both MCP client and MCP server: it registers your servers, filters tools per virtual key, brokers OAuth and checks every tool call before it executes.

  • Registry, Agent Mode and Code Mode on every plan
  • OAuth 2.1 with PKCE
  • Pinned tool definitions
MCP gateway illustration: tool icons and servers connected through one shielded gateway

Why MCP needs a gateway

The Model Context Protocol defines how clients discover and call tools. Who may call which tool, with which credentials, is left to you.

Tool sprawl

Every team connects its own MCP servers, so nobody can list which tools your agents can reach.

Credentials everywhere

API tokens are copied into agent configs and scripts, with no central place to scope or revoke them.

Tools that change

A server can change a tool’s description or schema after review, and clients keep calling it.

How the MCP gateway works

The MCP gateway shares one policy engine with the rest of the DeepintShield platform. See how a request flows from caller to tool.

Connect and register

Add servers over HTTP, SSE or STDIO; the gateway health-checks them and lists their tools.

Filter per key

Each virtual key sees only the tools its MCP configuration allows. An empty configuration means no tools.

Authorize each call

On Business and above, every tool call also needs an exact Allow from Agentic Security.

Record the result

Tool calls land in the logs, and agentic decisions go to the hash-chained ledger.

What a caller gets back

With Agentic Security on, the direct tool-execution endpoint returns one of three outcomes. JSON-RPC callers get the same verdict as an error result.

Allow

The tool runs

The key binding, the tool filter and the authorization check all allow the call, so the gateway executes it.

Deny

The call is refused

Returns 403 mcp_tool_authorization_denied. Auto-execute lists and OAuth grants cannot override it.

Require approval

The call waits

Returns 202 mcp_tool_approval_required. Write-like actions need a live approval tied to the argument digest.

MCP gateway features

Registry and health checks

Register HTTP, SSE and STDIO servers, with health checks every 10 seconds. STDIO is off in production by default.

Tool filtering

Client config, virtual key binding and authorization intersect. Request headers can narrow the tool list but never expand it.

OAuth 2.1 with PKCE

Strict OAuth 2.1 (MCP 2026-07-28) with PKCE S256, issuer binding and encrypted tokens bound to the workspace.

Delegated authentication

RFC 8693 token exchange. The caller’s token is never forwarded to the MCP server and never stored.

Agent Mode and Code Mode

Agent Mode runs tool loops for you; Code Mode calls tools from one sandboxed Starlark script, checking every nested call.

Safe outbound connections

Server URLs are validated before the gateway connects, and cloud metadata endpoints are refused unconditionally.

MCP servers per plan: Developer 1, Team 25, Business 50, Enterprise unlimited. The registry, Agent Mode and Code Mode are on every plan; per-call agentic authorization is on Business and above. See pricing.

Pinned tool definitions catch MCP server drift

A tool approval covers the tool’s name, description and argument schema. If an upstream MCP server changes any of them, the tool moves to Drifted and calls are denied. The console shows the approved and advertised digests, and an operator chooses Approve new definition to re-pin the tool.

  • Definition states: No pin yet, Matches, Drifted, Reverted
  • The server classifies actions and ignores caller-supplied classes
  • Once any agent is enrolled, ungoverned MCP is refused
MCP gateway tool pinning lifecycle: the approved tool definition’s digest is compared with the live one on every call; a match goes ahead, drift is denied until re-approved(opens the full-size image in a new tab)
Tool pinning and drift detection on every governed call.

An inventory of every governed tool

Export AIBOM produces a JSON inventory of governed tool contracts, so security and audit teams can see which tools agents may use and which definitions were approved. On Business and above, the export is signed when a signing key is configured; otherwise it is clearly marked unsigned. Read how agentic AI security uses the same contracts.

MCP gateway agent inventory in the DeepintShield console: governed agents with their framework, tools, source and status(opens the full-size image in a new tab)
Export AIBOM from the agent inventory. The file is signed only when a signing key is configured.

Point your agents at one governed endpoint

Agents and frameworks reach MCP tools through the gateway with a virtual key; anonymous access is refused, and each key gets its own filtered catalogue. The /mcp gateway URL accepts JSON-RPC 2.0 and SSE and needs gateway version v1.4.0-prerelease1 or later.

Endpoints and responses
# Governed MCP entry points
/mcp                        # JSON-RPC 2.0 and SSE
POST /v1/mcp/tool/execute   # direct tool execution

# Responses from the authorization check
403 mcp_tool_authorization_denied
202 mcp_tool_approval_required
503 mcp_tool_authorization_unavailable

Mapped to agent and LLM risk frameworks

Mapped to

  • OWASP Agentic Top 10
  • OWASP LLM Top 10
  • MITRE ATLAS

Mapping supports your own assessment; it does not replace it.

Further reading on MCP security

MCP gateway questions

What is an MCP gateway?

An MCP gateway is a single control point between AI clients and Model Context Protocol servers. It handles discovery, authentication and policy for every tool call, so tools are not wired directly into each agent. Read the MCP overview in the docs.

Is MCP secure by default?

No. The protocol describes how to discover and call tools, but access control and credential handling are up to you. DeepintShield adds per-key tool filtering, OAuth 2.1 and, on Business and above, an authorization check on every call.

How are MCP credentials handled?

Credentials stay on the gateway, not in your agents. OAuth tokens are encrypted and bound to the workspace, and delegated sign-in never forwards the caller's token to the MCP server. Set DEEPINTSHIELD_ENCRYPTION_KEY so that stored server secrets are encrypted.

Which MCP transports are supported?

STDIO, HTTP and SSE are supported. STDIO servers are off in production by default, and outbound URLs are validated before the gateway connects.

How many MCP servers can I connect?

It depends on your plan: Developer 1, Team 25, Business 50 and Enterprise unlimited, counted across your organisation. See the plan comparison.

Govern every MCP tool call

Connect your first MCP server on the free Developer plan, or book a demo to see per-call authorization.

Start For Free