Skip to content

INTEGRATIONS.md

Latest commit

 

History

History
115 lines (88 loc) · 3.64 KB

INTEGRATIONS.md

File metadata and controls

115 lines (88 loc) · 3.64 KB
Raw

Integrations

Recipes for putting Prompt Anonymizer where your LLM traffic already flows. Most of these build on the OpenAI-compatible proxy (@prompt-anonymizer/proxy) — anything that speaks the OpenAI API can be pointed at it with a single base-URL change; PII is masked on the way out and labels are restored in the reply (streaming included).

Start the proxy once (see the README quickstart):

npx @prompt-anonymizer/proxy
# → http://127.0.0.1:8787/v1  (admin GUI: http://127.0.0.1:8787/admin/)

Anything that reads OPENAI_BASE_URL

The OpenAI SDKs (Python / JS), llm, aider, Continue, and most tools honor the standard environment variables:

export OPENAI_BASE_URL=http://127.0.0.1:8787/v1
export OPENAI_API_KEY=sk-...        # forwarded upstream unchanged

LiteLLM

Point the model's api_base at the proxy — per model, so you can route only the vendors you don't want to see PII:

# litellm config.yaml
model_list:
  - model_name: gpt-4o-masked
    litellm_params:
      model: openai/gpt-4o
      api_base: http://127.0.0.1:8787/v1
      api_key: os.environ/OPENAI_API_KEY

Or in code: litellm.completion(model="openai/gpt-4o", api_base="http://127.0.0.1:8787/v1", ...).

OpenWebUI

Add the proxy as an OpenAI-compatible connection: Admin Panel → Settings → Connections → OpenAI API → set the URL to http://127.0.0.1:8787/v1 (use http://host.docker.internal:8787/v1 when OpenWebUI runs in Docker and the proxy on the host). Every chat through that connection is anonymized before it leaves the machine; the reply comes back restored, so users see real names while the upstream vendor never does.

MCP clients (Claude Desktop / Claude Code / Cursor)

The MCP server exposes anonymize / deanonymize / scan as tools — see the README quickstart and web/packages/mcp/README.md. Claude Code:

claude mcp add prompt-anonymizer -- npx -y @prompt-anonymizer/mcp

Typical agent uses: scan a repo before pushing; anonymize a local file (file: input) so its PII never enters the model context; deanonymize with output_file to restore a reply straight to disk.

Git hooks and CI

pre-commit framework (Python CLI) — full recipe in the README:

repos:
  - repo: https://github.com/akazah/prompt-anonymizer
    rev: v0.3.0
    hooks:
      - id: prompt-anonymizer-scan

husky + lint-staged (Node CLI):

{ "lint-staged": { "*": "prompt-anonymizer scan" } }

GitHub Actions job that fails a PR when tracked text files contain PII (findings show file:line:col and entity type only — logs stay PII-free):

jobs:
  pii-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v7
      - uses: actions/setup-python@v6
        with: { python-version: "3.13" }
      - run: pip install prompt-anonymizer==0.3.0
      - run: git ls-files -z '*.md' '*.txt' | xargs -0 prompt-anonymizer scan

Python apps (library, no proxy)

When you control the code, calling the library directly gives you the mapping for the round-trip — see the README quickstart. The same applies to JS/TS apps with @prompt-anonymizer/core or the React/Vue bindings.

Notes

  • The proxy keeps mappings in memory per request and binds to 127.0.0.1 by default. Detection there is regex-first with opt-in NER — configure entity types and deny/allow lists in the admin GUI.
  • Detection is best-effort everywhere; treat these integrations as a safety net, not a guarantee (see README limitations).