A pluggable, language-agnostic filter framework for Model Context Protocol (MCP) traffic. Each filter is a tiny HTTP service that decides whether a tool call should be allowed, blocked, modified, or truncated.

MCP gives agents direct access to tools and data sources. That's powerful and dangerous in equal measure. Filters give you a place to enforce guardrails — block destructive SQL, redact PII before it reaches the model, throttle abusive agents, cap response sizes, redact columns — without modifying the upstream MCP server or the agent runtime.

                    pre-request                       post-response
   Client ─▶ MCP Gateway ─────────▶ Engine ─▶ filter ─┐
                │     ◀────────────────── allow/block ┘
                │
                ▼ (forward)
            MCP Server
                │
                ▼
   Client ◀ MCP Gateway ◀──── Engine ─▶ filter ─┐
                                ◀───── allow/truncate/modify ┘

• An MCP gateway sits in front of the upstream MCP server.
• For every tools/call, the gateway asks an engine "what should I do for target X at execution_point Y?".
• The engine looks up the policies assigned to that target and, for each filter in those policies, calls POST {filter_url}/filter.
• Filters are completely independent processes. They can be written in any language. They share no state with the gateway or the engine.

A reference engine and gateway live in the parent repo (secure-ai-plane-x/mcp-filter-gateway). This sub-project ships only the filter implementations and the contract.

Every filter MUST expose two HTTP endpoints:

Request body:

{
  "config":    { "max_rows": 1000 },
  "arguments": {
    "sql": "SELECT * FROM employees",
    "response_count": 8000,
    "response_rows": ["row-1", "row-2", "..."]
  },
  "metadata": {
    "agent_id":   "demo_agent",
    "agent_name": "demo_agent",
    "member_id":  "f5d9e4ec-…",
    "project_id": "b0000000-…",
    "tool_name":  "run_sql"
  }
}

Response body:

{
  "action": "allow | block | truncate | modify",
  "reason": "human-readable string surfaced to the user and audit log",
  "modified_response": { "...optional..." },
  "original_count":    8000,
  "truncated_to":      1000
}

• allow — let the request through unchanged.
• block — reject. The gateway returns an MCP error using reason.
• truncate — return a smaller payload (modified_response) and an audit notice. Used by post-response filters like Row Limiter.
• modify — replace the payload with modified_response. Used by redactors and maskers.

Filters MUST be idempotent and side-effect free with respect to the agent's behaviour (with the obvious exception of audit / rate limiter state). Filter execution time should stay under 50ms p95; the engine treats slow filters as allow to fail-open.

Returns 200 { "status": "healthy", "filter": "<slug>", "version": "..." }. Used by orchestrators to gate traffic.

Cross-cutting guards under common/ (Apache 2.0):

Integration-specific guards under provider/<vendor>/data-guard/ — 54 providers today (AWS, GitHub, Stripe, Snowflake, Slack, PostgreSQL, Notion, Terraform, and more). See provider/README.md and provider/manifest.json for the full list and ports (6301–6354).

Run node provider/scripts/generate-providers.js after editing provider/manifest.json to scaffold new vendors.

cd common/sql-guard
npm install
PORT=6001 node index.js

Or bring up all filters with Docker:

docker compose up

Then register the filter against an engine:

INSERT INTO ablv_mcp_filters (slug, name, filter_url, execution_point, ...)
VALUES ('sql-guard', 'SQL Guard', 'http://sql-guard:6001/filter',
        'pre-request', ...);

See docs/ADDING_FILTERS.md for a step-by-step walkthrough of writing your own filter.

mcp-filters/
├── README.md              # this file
├── LICENSE                # Apache 2.0
├── docker-compose.yml     # common + provider filters
├── docs/
│   ├── CONTRACT.md        # full HTTP contract + JSON schema
│   ├── ARCHITECTURE.md    # how filters compose with gateway + engine
│   └── ADDING_FILTERS.md  # write-your-own walkthrough
├── common/                # cross-cutting filters (any MCP server)
│   ├── sql-guard/
│   ├── pii-redactor/
│   ├── row-limiter/
│   ├── schema-validator/
│   ├── rate-limiter/
│   └── field-masker/
├── provider/              # vendor-specific data guards (54 providers)
│   ├── manifest.json
│   ├── scripts/generate-providers.js
│   ├── _lib/
│   └── <vendor>/data-guard/
└── tests/
    ├── contract.js        # HTTP contract for every filter
    └── behaviour.js       # decision-logic regression tests