SafeNode MCP Gateway

SafeNode MCP Gateway

An MCP proxy that enforces policy on every tool call. The agent cannot route around it.

npx safenode-mcp-gateway

Why a proxy

An MCP server that exposes an evaluate_action tool is theatre. The model can choose not to call it, which means it enforces nothing.

This is a proxy. It sits between your MCP client (Claude Desktop, Claude Code, Cursor, any MCP host) and the MCP servers you already use. tools/list passes straight through, so the model sees exactly the tools it saw before. Every tools/call is evaluated against SafeNode *first*. A denial means the call never reaches the downstream server at all.

Claude Desktop ──► safenode-mcp-gateway ──► filesystem server

│ github server ▼ postgres server SafeNode API allow/warn/review/deny

Your agent needs no code changes. It does not know the gateway is there.

Get an API key — free tier, no card · Docs · Python SDK

---

Setup

1. Get an API key

Free at safenode.tech. Keys look like sn_....

2. Write a config

safenode-gateway.json — take the servers straight out of your existing MCP client config:

{

"failMode": "fail_closed", "payloadMode": "redacted", "logFile": "./safenode-decisions.jsonl", "servers": [ { "name": "filesystem", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects"] } ], "tools": { "read_file": { "failMode": "fail_open" }, "write_file": { "payloadMode": "metadata_only" } } }

Do not put your API key in this file. Set SAFENODE_API_KEY in the environment — it takes precedence over the config, so the file stays safe to commit.

3. See what would be sent, before sending anything

SAFENODE_API_KEY=sn_... npx safenode-mcp-gateway --dry-run

Prints the exact JSON body that an evaluation would POST. No network call. This is the fastest way to satisfy yourself about what leaves your machine.

4. Point your MCP client at the gateway

Claude Desktop — claude_desktop_config.json:

{

"mcpServers": { "safenode": { "command": "npx", "args": ["-y", "safenode-mcp-gateway", "--config", "/absolute/path/to/safenode-gateway.json"], "env": { "SAFENODE_API_KEY": "sn_..." } } } }

Claude Code — .mcp.json in your project:

{

"mcpServers": { "safenode": { "command": "npx", "args": ["-y", "safenode-mcp-gateway", "--config", "./safenode-gateway.json"], "env": { "SAFENODE_API_KEY": "sn_..." } } } }

Cursor — ~/.cursor/mcp.json, same shape as Claude Desktop.

Replace your existing server entries with this one. The gateway spawns them itself, so listing them in both places would run each server twice.

---

What each decision does

Denials come back as MCP tool *errors*, not transport failures, so the model sees why and can try something else instead of the host surfacing an opaque crash.

Fail behaviour

failMode decides what happens when SafeNode itself is unreachable.

The gateway defaults to `fail_closed`, unlike the SafeNode SDKs, which default to `fail_open`. That difference is deliberate. An SDK wraps a developer's own code, where taking production down during a SafeNode outage loses the user forever. The gateway fronts arbitrary MCP tools it knows nothing about — filesystem writes, shell commands, payments — and silently letting all of that through the moment SafeNode is unreachable defeats the point of installing it.

Loosen it per tool for the read-only ones:

"tools": {

"read_file": { "failMode": "fail_open" }, "list_files": { "failMode": "fail_open" } }

Degraded decisions are logged to stderr and marked "degraded": true with a null traceId in the decision log, so they can never be counted as real policy decisions.

What data leaves your machine

For every tool call, the gateway sends: the tool name, the downstream server name, a session id, any static context from your config, and the tool arguments as the payload.

payloadMode controls the arguments:

Redaction covers emails, Luhn-validated credit cards, US SSNs, provider API key prefixes (sk-, ghp_, xoxb-, AKIA, AIza), bearer tokens, and PEM private key blocks. Values become [REDACTED: ].

Why redaction counts are always sent

Alongside the redacted payload, the gateway sends counts of what it removed:

"safenode_redactions": { "email": 2, "credit_card": 1 }

This is load-bearing. SafeNode's server-side sensitive_data rule matches patterns against payload *values*. If the gateway scrubbed those values and said nothing, a policy of "deny any action containing a card number" would silently start passing — the client-side privacy feature would have disabled the server-side security control. Reporting counts closes that: policy can act on the *presence* of a card number without ever receiving one.

If you use sensitive_data with patterns, pair it with a redaction_metadata rule.

These counts are self-reported. They raise the floor for an honest client; they are not a defence against a hostile one.

No telemetry

The gateway talks to exactly two things: the downstream MCP servers you configured, and the SafeNode API. No analytics, no phone-home, no postinstall scripts.

Local decision log

Set logFile to get a JSONL record of every decision, independent of the server-side audit trail — so you still have evidence of what your agent tried to do when the network was down:

{"timestamp":"2026-08-08T04:12:09.412Z","server":"filesystem","tool":"write_file","decision":"deny","degraded":false,"traceId":"9f1c…","reasons":["Path is outside the approved workspace."],"forwarded":false,"durationMs":143,"sessionId":"a3f1…"}

Configuration reference

failMode, payloadMode and context can be set globally, per server, or per tool. Most specific wins.

tools. .bypass: true skips evaluation entirely for one tool. It does what it says — the call is forwarded with no policy check at all.

Tool name collisions

If two downstream servers export the same tool name, the gateway prefixes every tool with __ and logs a warning. Silently picking a winner would apply one server's policy to another server's tool, which is a security bug rather than a cosmetic one.

What this is not

• Not a sandbox. It decides whether a call is forwarded. It does not contain the downstream server, restrict syscalls, or limit what that server can do once the call reaches it.
• Not a prompt-injection detector. It evaluates the *action*, not the reasoning that produced it. If your agent has been talked into deleting a table, this can stop the delete — it will not tell you the agent was manipulated.
• Not offline. Every tool call costs a round trip to the SafeNode API. There is no local policy evaluation.
• Not a substitute for scoping the underlying servers. Still pass a filesystem server the narrowest root directory that works.
• Not free of side effects. Every evaluation is recorded server-side and counts against your monthly quota.

Latency

Each gated tool call adds one round trip to the SafeNode API, with a 5000ms budget by default. Bypassed tools add nothing.

A measured p99 for the evaluate endpoint is not published yet, because it has not been measured under realistic load. When it has been, it will go here and in the docs rather than being estimated.

Docker

docker build -t safenode-mcp-gateway .

Running it with no arguments starts against a bundled demo config that proxies an included echo server, so you can see the proxy work without any setup:

docker run -i --rm safenode-mcp-gateway

The demo key is a placeholder, so tools/list passes through while any real tool call is denied by the fail_closed policy. For real use, mount your own config and supply a key:

docker run -i --rm \

-e SAFENODE_API_KEY=sn_... \ -v /path/to/safenode-gateway.json:/app/config.json \ safenode-mcp-gateway --config /app/config.json

-i is required. MCP speaks JSON-RPC over stdio, so without stdin attached the gateway has nothing to talk to.

Agent skill

SKILL.md teaches an agent to set the gateway up for you, moving your existing servers behind it, and how to behave when a call is blocked or held for review. Copy it into a folder in your agent's skills directory (for Claude Code, ~/.claude/skills/safenode-mcp-gateway/). The policy-authoring companion is in skill/safenode-policies/.

Requirements

Node 18+. One runtime dependency: @modelcontextprotocol/sdk.

Contributing

See CONTRIBUTING.md. Bug reports about enforcement gaps — anything that reaches a downstream server when it should have been blocked — are the most valuable thing you can file.

Security

See SECURITY.md. Please do not open public issues for vulnerabilities.

License

MIT. See LICENSE.