Claude Code read my .env and sent the API key to Anthropic. I built a local check that withholds it.
retained text
And now the solution works for every tool call, not just your .envs. Think log traces, sensitive emails, not just that one file.
When Claude Code reads a file or runs a command, the output goes into its next request to Anthropic. If that output has an API key or a password in it (an `.env`, a config file, a log, `printenv`), the key goes too.
`permissions.deny` on `Read(.env)` helps for files you know about. It doesn't catch a token in a log, a test fixture, or command output.
So I put a small local model in the middle. Before each request leaves, it scans the new tool output. If it spots a secret, Claude gets **"Tool output withheld"** plus a reason instead of the content, and can try another approach (for example, reading a narrower range). Nothing is sent to anyone but your local model.
**Setup (about 5 minutes, macOS/Linux):**
# 1. local scanner model (~3.3 GB)
llama-server -hf bartowski/Qwen2.5-3B-Instruct-GGUF:Q8_0 --alias local-pii \
--host 127.0.0.1 --port 8081 -c 8192 -np 1 --jinja --no-context-shift
# 2. the proxy + plugin
curl -fsSL https://torana.sh/install.sh | sh
torana start --port 8143
torana plugin install pii
torana open # add the local model as a provider, bind it to the pii plugin
# 3. run Claude Code through it
ANTHROPIC_BASE_URL="$(torana endpoint anthropic)" claude
Your normal Claude login and billing stay the same.
**How well does a small model do this?** 60 files with planted (fake) secrets, 60 harmless ones:
|Scanner|Secrets caught|Harmless files blocked|
|:-|:-|:-|
|Qwen2.5 3B|60 / 60|18 / 60|
|privacy-filter|58 / 60|9 / 60|
|Regex only|24 / 60|0 / 60|
**What I learned building it:**
* **Withhold, but say why.** The plugin replaces the whole result with a short reason, so Claude can change approach (for example, read a narrower range) instead of just retrying.
* **Valid JSON isn't a correct answer.** Qwen3 0.6B answered every case in clean JSON and missed all 60 secrets. Most sub-1B models couldn't even keep the format.
* **Prompt wording matters a lot.** Two rewordings of the same prompt for Qwen2.5 3B caught 37 and 49 of 52 secrets in real code.
* **Specialised detectors are fast but jumpy.** SecretMasker takes about 11 ms, but it mistakes random hashes for keys.
**Honest limits:**
* It adds about a second per tool result.
* It has false alarms. Expect the odd harmless file to be withheld.
* It's an extra layer, not a guarantee.
Full write-up with 15 models and detectors compared, plus the step-by-step setup (disclosure: I built Torana, it's open source and free): [https://torana.sh/blog/local-models-tool-output/](https://torana.sh/blog/local-models-tool-output/?utm_source=reddit&utm_medium=social&utm_content=claudecode-pii)
How are you keeping secrets out of Claude Code's context today? Deny rules, hooks, something else?document fetched 2026-10-13 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1x4a6kg/claude_code_read_my_env_and_sent_the_api_key_to