Prompt Cleaner (MCP Server)

TypeScript MCP server exposing a prompt cleaning tool and health checks. All prompts route through cleaner, with secret redaction, structured schemas, and client-friendly output normalization.

Features

• Tools
  • health-ping: liveness probe returning { ok: true }.
  • cleaner: clean a raw prompt; returns structured JSON with retouched string, notes, openQuestions, risks, and redactions.
• Secret redaction: Sensitive patterns are scrubbed from logs and outputs in src/redact.ts.
• Output normalization: src/server.ts converts content with type: "json" to plain text for clients that reject JSON content types.
• Configurable: LLM base URL, API key, model, timeout, log level; optional local-only enforcement.
• Deterministic model policy: Single model via LLM_MODEL; no dynamic model selection/listing by default.

Requirements

• Node.js >= 20

Install & Build

npm install
npm run build

Run

• Dev (stdio server):

npm run dev

Inspector (Debugging)

Use the MCP Inspector to exercise tools over stdio:

npm run inspect

Environment

Configure via .env or environment variables:

• LLM_API_BASE (string, default http://localhost:1234/v1): OpenAI-compatible base URL.
• LLM_API_KEY (string, optional): Bearer token for the API.
• LLM_MODEL (string, default open/ai-gpt-oss-20b): Model identifier sent to the API.
• LLM_TIMEOUT_MS (number, default 60000): Request timeout.
• LOG_LEVEL (error|warn|info|debug, default info): Log verbosity (logs JSON to stderr).
• ENFORCE_LOCAL_API (true|false, default false): If true, only allow localhost APIs.
• LLM_MAX_RETRIES (number, default 1): Retry count for retryable HTTP/network errors.
• RETOUCH_CONTENT_MAX_RETRIES (number, default 1): Retries when the cleaner returns non-JSON content.
• LLM_BACKOFF_MS (number, default 250): Initial backoff delay in milliseconds.
• LLM_BACKOFF_JITTER (0..1, default 0.2): Jitter factor applied to backoff.

Example .env:

LLM_API_BASE=http://localhost:1234/v1
LLM_MODEL=open/ai-gpt-oss-20b
LLM_API_KEY=sk-xxxxx
LLM_TIMEOUT_MS=60000
LOG_LEVEL=info
ENFORCE_LOCAL_API=false
LLM_MAX_RETRIES=1
RETOUCH_CONTENT_MAX_RETRIES=1
LLM_BACKOFF_MS=250
LLM_BACKOFF_JITTER=0.2

Tools (API Contracts)

All tools follow MCP Tool semantics. Content is returned as [{ type: "json", json: <payload> }] and normalized to type: "text" by the server for clients that require it.

• health-ping

  • Input: {}
  • Output: { ok: true }

• cleaner

  • Input: { prompt: string, mode?: "code"|"general", temperature?: number }
  • Output: { retouched: string, notes?: string[], openQuestions?: string[], risks?: string[], redactions?: ["[REDACTED]"][] }
  • Behavior: Applies a system prompt from prompts/cleaner.md, calls the configured LLM, extracts first JSON object, validates with Zod, and redacts secrets.

• sanitize-text (alias of cleaner)

  • Same input/output schema and behavior as cleaner. Exposed for agents that keyword-match on “sanitize”, “PII”, or “redact”.

• normalize-prompt (alias of cleaner)

  • Same input/output schema and behavior as cleaner. Exposed for agents that keyword-match on “normalize”, “format”, or “preprocess”.

Per-call API key override

src/llm.ts accepts apiKey in options for per-call overrides; falls back to LLM_API_KEY.

Project Structure

• src/server.ts: MCP server wiring, tool listing/calls, output normalization, logging.
• src/tools.ts: Tool registry and dispatch.
• src/cleaner.ts: Cleaner pipeline and JSON extraction/validation.
• src/llm.ts: LLM client with timeout, retry, and error normalization.
• src/redact.ts: Secret redaction utilities.
• src/config.ts: Environment configuration and validation.
• test/*.test.ts: Vitest suite covering tools, shapes, cleaner, and health.

Testing

npm test

Design decisions

• Single-model policy: Uses LLM_MODEL from environment; no model listing/selection tool to keep behavior deterministic and reduce surface area.
• Output normalization: src/server.ts converts json content to text for clients that reject JSON.
• Secret redaction: src/redact.ts scrubs sensitive tokens from logs and outputs.

Troubleshooting

• LLM timeout: Increase LLM_TIMEOUT_MS; check network reachability to LLM_API_BASE.
• Non-JSON from cleaner: Retries up to RETOUCH_CONTENT_MAX_RETRIES. If persistent, reduce temperature or ensure the configured model adheres to the output contract.
• HTTP 5xx from LLM: Automatic retries up to LLM_MAX_RETRIES with exponential backoff (LLM_BACKOFF_MS, LLM_BACKOFF_JITTER).
• Local API enforcement error: If ENFORCE_LOCAL_API=true, LLM_API_BASE must point to localhost.
• Secrets in logs/outputs: Redaction runs automatically; if you see leaked tokens, update patterns in src/redact.ts.

Windsurf (example)

Add an MCP server in Windsurf settings, pointing to the built stdio server:

{
  "mcpServers": {
    "prompt-cleaner": {
      "command": "node",
      "args": ["/absolute/path/to/prompt-cleaner/dist/server.js"],
      "env": {
        "LLM_API_BASE": "http://localhost:1234/v1",
        "LLM_API_KEY": "sk-xxxxx",
        "LLM_MODEL": "open/ai-gpt-oss-20b",
        "LLM_TIMEOUT_MS": "60000",
        "LOG_LEVEL": "info",
        "ENFORCE_LOCAL_API": "false",
        "LLM_MAX_RETRIES": "1",
        "RETOUCH_CONTENT_MAX_RETRIES": "1",
        "LLM_BACKOFF_MS": "250",
        "LLM_BACKOFF_JITTER": "0.2"
      }
    }
  }
}

Usage:

• In a chat, ask the agent to use cleaner with your raw prompt.
• Or invoke tools from the agent UI if exposed by your setup.

LLM API compatibility

• Works with OpenAI-compatible Chat Completions APIs (e.g., LM Studio local server) that expose /v1/chat/completions.
• Configure via LLM_API_BASE and optional LLM_API_KEY. Use ENFORCE_LOCAL_API=true to restrict to localhost for development.
• Set LLM_MODEL to the provider-specific model identifier. This server follows a single-model policy for determinism and reproducibility.
• Providers must return valid JSON; the cleaner includes limited retries when content is not strictly JSON.

Links

• Model Context Protocol (spec): https://modelcontextprotocol.io
• Cleaner system prompt: prompts/cleaner.md

Notes

• Logs are emitted to stderr as JSON lines to avoid interfering with MCP stdio.
• Some clients reject json content types; this server normalizes them to text automatically.

Security

• Secrets are scrubbed by src/redact.ts from logs and cleaner outputs.
• ENFORCE_LOCAL_API=true restricts usage to local API endpoints.