killswitch-ai 0.1.5

Local LLM egress control for Python developers — prevents API keys, secrets, .env files, and sensitive payloads from being sent to LLM APIs

killswitch-ai

Local LLM egress control for Python developers. Stops API keys, secrets, .env files, and sensitive payloads before they reach OpenAI, Anthropic, or any LLM API — without leaving your machine.

---

Install

pip install killswitch-ai

With provider extras:

pip install "killswitch-ai[openai]"      # OpenAI
pip install "killswitch-ai[anthropic]"   # Anthropic
pip install "killswitch-ai[all]"         # Both

---

Quick start

One-liner

import killswitch
killswitch.install()

from openai import OpenAI
client = OpenAI()  # now protected — every request is scanned first

Explicit wrapper

from openai import OpenAI
from killswitch.openai import GuardedOpenAI

client = GuardedOpenAI(OpenAI())

Anthropic

from anthropic import Anthropic
from killswitch.anthropic import GuardedAnthropic

client = GuardedAnthropic(Anthropic())

Scan text directly

import killswitch

result = killswitch.scan("my API_KEY is sk-proj-abc123...")
for finding in result.findings:
    print(finding.severity, finding.description)

---

How it works

Your code  →  killswitch  →  LLM API
               ↓
           scan prompt
               ↓
         finding detected?
          ┌──────────────────────────────┐
          │  kill        →  raise error  │
          │  pause       →  ask you      │
          │  redact      →  [REDACTED]   │
          │  report_only →  log + allow  │
          └──────────────────────────────┘

Everything runs locally. No prompts, responses, file contents, or secret values leave your machine for inspection.

---

What gets detected

Layer	Examples
Prohibited terms	API_KEY, SECRET_KEY, CONFIDENTIAL, custom terms
Secret patterns	OpenAI keys (sk-proj-…), Anthropic keys, AWS keys, GitHub tokens, Stripe keys, private key PEM blocks, JWT tokens, database URLs
Entropy detection	Long random-looking strings that might be secrets
Sensitive file paths	.env, *.pem, id_rsa, credentials.json, kubeconfig
Structured payloads	Recursively scans every field in nested JSON

---

Modes

Mode	What happens
pause	Stops and asks you what to do (default)
kill	Blocks the request automatically
redact	Replaces secrets with [REDACTED_X] and continues
report_only	Allows the request but logs a finding

Change mode anytime:

killswitch mode kill
killswitch mode pause
killswitch mode redact
killswitch mode report-only

---

Set up

killswitch init

Runs an interactive wizard: choose a mode, pick data categories to protect, optionally enable weekly email summaries. Writes killswitch.yml to your project.

Then verify installation:

killswitch test

---

CLI

killswitch init                    # Interactive setup wizard
killswitch test                    # Verify the scanner is working
killswitch status                  # Show current mode and recent stats
killswitch scan "text to scan"     # Test the scanner on text
killswitch menu                    # Open the interactive report menu
killswitch mode kill               # Change active mode
killswitch logs --latest           # View latest session findings
killswitch event KAI-E-...         # Look up a specific event
killswitch finding KAI-F-...       # Look up a specific finding
killswitch report                  # Print the weekly summary
killswitch report --send           # Send weekly report by email
killswitch email --on              # Enable email reports
killswitch email --off             # Disable email reports

---

Configuration (killswitch.yml)

mode:
  default_action: pause            # kill | pause | redact | report_only

detection:
  prohibited_terms:
    - "API_KEY"
    - "SECRET_KEY"
    - "PRIVATE KEY"
    - "MY_CUSTOM_TERM"
  entropy_detection:
    enabled: true
    min_length: 24
    threshold: 4.2

actions:
  private_key: kill
  openai_key: kill
  anthropic_key: kill
  database_url: pause
  prohibited_term: pause
  high_entropy_string: report_only

logging:
  enabled: true
  log_dir: .killswitch
  store_raw_payloads: false        # never stores prompt text
  store_secret_values: false       # never stores secret values

email:
  enabled: false
  address: ""
  frequency: weekly

---

Local logs

Every scan writes an audit event to .killswitch/sessions/. Raw prompts and secret values are never stored.

Stored: event ID, timestamp, provider, operation, decision (blocked / allowed / redacted), finding type, severity, recommendations.

Never stored: prompt text, LLM responses, secret values, API keys, source code, file contents.

---

Privacy

• All scanning runs locally — nothing leaves your machine for inspection
• No prompts or responses are uploaded
• No secret values are stored in logs or telemetry
• Anonymous opt-in telemetry reports only aggregate counts (scans, blocks, finding categories) — disabled by default

---

Links

• Website: killswitch-ai.com
• PyPI: pypi.org/project/killswitch-ai
• Issues: github.com/killswitch-ai/killswitch/issues

---

License

MIT