mcp-firewall

Policy firewall for MCP servers.

mcp-firewall sits between an AI client and downstream MCP servers, evaluates policy before side effects, and returns explainable decisions.

If you are exposing MCP tools to autonomous or semi-autonomous agents, this gives you a practical control layer with a short setup path.

The hook in one minute

Most MCP integrations start with direct agent-to-tool connectivity. That is fast, but operationally risky:

broad tool reach by default
weak visibility when something is blocked or allowed
logs that may contain sensitive values

mcp-firewall changes the default path:

evaluate allow/deny/confirmation rules first
enforce path/repo/domain scope constraints
redact common secret patterns in audit logs
return policy-linked deny responses with remediation hints

Proven local result (reproducible)

npm install
npm run qa:local

qa:local runs CI gates, boots a deterministic mock MCP server, verifies one safe call is allowed, and verifies one risky call is denied.

Representative output from a successful run:

[1/4] Run CI quality gates
... 19 tests passed ...
[2/4] Start mcp-firewall mock server
[3/4] Wait for health endpoint
[4/4] Run smoke test
Health status: ok
tools/list returned 3 tools
Safe call succeeded with tool: mockfs__read_file
Blocked call confirmed. policyId: policy.tools.deny

Who this is for

teams connecting internal agents to filesystem, git, or network-capable MCP tools
maintainers who need clear policy behavior for demos, governance, and incident review
developers who want a thin control plane, not a full agent framework

Architecture

flowchart LR
  A[AI Client or Agent] --> B[mcp-firewall]
  B --> C[Policy Engine]
  C --> D{Allow?}
  D -- No --> E[Explainable Deny Response]
  D -- Yes --> F[Forwarder]
  F --> G[Downstream MCP Servers]
  B --> H[Redaction Layer]
  H --> I[Structured Audit Logs]

Quickstart

Install and build.

npm install
npm run build

Copy a preset and edit it.

Windows PowerShell:

Copy-Item presets\local-safe.json config.json

macOS/Linux:

cp presets/local-safe.json config.json

Start firewall.

npx mcp-firewall start --config ./config.json

Inspect merged MCP surface.

npx mcp-firewall list --config ./config.json

Example deny response

{
  "jsonrpc": "2.0",
  "id": 3,
  "error": {
    "code": -32010,
    "message": "Tool 'mockfs__delete_file' is denied by policy.",
    "data": {
      "code": "tool_denied",
      "policyId": "policy.tools.deny",
      "matchedRule": "*delete*",
      "remediationHint": "Remove or narrow this deny rule if tool access is intended."
    }
  }
}

Policy model

allow: explicit allowlist patterns
deny: explicit denylist patterns
requireConfirmation: sensitive operations that require user confirmation upstream
scope: path/repo/domain constraints
redaction: log masking for common secret patterns

Example policy skeleton:

{
  "policy": {
    "mode": "hybrid",
    "tools": {
      "allow": ["*read*", "*list*", "search*"],
      "deny": ["*delete*", "*push*", "*write*"],
      "requireConfirmation": ["*exec*", "*shell*"]
    },
    "scope": {
      "paths": {
        "allow": ["."],
        "deny": [".git", ".env", "**/secrets/**"]
      },
      "repos": {
        "allow": ["your-org/*"],
        "deny": []
      },
      "domains": {
        "allow": [],
        "deny": ["*"]
      }
    },
    "redaction": {
      "enabled": true,
      "extraPatterns": []
    }
  }
}

Design rationale (decision -> reason -> source)

| Decision | Why this choice | Source | | --- | --- | --- | | Policy decision before forwarding | Prevent side effects when a request violates policy; fail early and explicitly. | [1], [2] | | JSON-RPC strictness (jsonrpc: "2.0") | Remove protocol ambiguity and keep client/server behavior deterministic. | [2] | | Structured audit logging | Make incident review and policy tuning practical in production. | [3] | | Secret redaction in logs | Lower accidental credential exposure during debugging and operations. | [4] | | Environment variable interpolation in config | Keep secrets out of committed config and align with deployment practice. | [5] |

Expanded rationale and tradeoffs live in docs/RATIONALE.md.

Included presets

presets/local-safe.json
presets/read-only.json
presets/github-safe.json

Commands

# Lint
npm run lint

# Build distributables
npm run build

# Run strict CI gates (typecheck + tests + build)
npm run qa:ci

# End-to-end local verification
npm run qa:local

# Smoke test only
npm run smoke:test

# Formatting checks for code/config files
npm run format:check

# Optional PowerShell variants
npm run qa:local:ps1
npm run smoke:test:ps1

Supported MCP methods

tools/list
tools/call
resources/list
resources/read
prompts/list
prompts/get

Limits and non-goals

not a replacement for host hardening, sandboxing, or IAM
not an enterprise auth platform in v0.1.x
not a guarantee of tool safety; it is a policy and observability control layer

Documentation

References

[1] Model Context Protocol documentation, https://modelcontextprotocol.io/

[2] JSON-RPC 2.0 Specification, https://www.jsonrpc.org/specification

[3] OWASP Logging Cheat Sheet, https://cheatsheetseries.owasp.org/cheatsheets/Logging_Cheat_Sheet.html

[4] OWASP Secrets Management Cheat Sheet, https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html

[5] The Twelve-Factor App, Config, https://12factor.net/config

MCP Servers