Your coding agent just read your API keys. I built a local check so they don't reach the model

Description and use

Torana’s pii plugin scans tool output with a local model to block API keys and passwords before they reach a hosted LLM.

Problem

Coding agents can inadvertently expose API keys or passwords to hosted LLMs when they include file contents in requests.

Outcome

The local scanner blocks 60 of 60 sensitive outputs while allowing most harmless outputs, with some false positives and false negatives depending on model and prompt.

Workflow

• 1. Agent reads file or runs tool.
• 2. Output sent to Torana’s pii plugin.
• 3. Plugin forwards output to a local model (e.g., Qwen2.5‑3B‑Instruct).
• 4. If model flags API key/password/PII, plugin replaces result with a short explanation.
• 5. Agent receives sanitized response and can retry with narrower scope.

Evidence

I built it as a plugin for Torana, an open‑source, local‑first proxy that sits between your coding agent and its model provider.

Source

If the model flags an API key, password or private personal data, the plugin replaces the whole tool result with a short explanation.

Source

I compared seven local chat models and several specialised detectors on 60 sensitive and 60 harmless synthetic examples.

Source

Scanning adds about a second per tool result, and large outputs that don't fit the model's context are blocked as scan failures.

Source