Docs & tutorial

Get your agent self-healing in minutes

SentinelProxy watches your agent's traffic, catches what a firewall only blocks, and hands you the patched system prompt. Here's how to wire it up and what every piece does.

Quickstart

Create a project & grab your key

In the dashboard, create a project — it mints a dedicated API key. Open its Connect tab to find the key and copy-paste snippets for every platform.

Connect your agent

Install the plugin for your stack (n8n, LangChain, LiteLLM, Dify) or fire the universal webhook after each agent turn. We return 200 instantly and scrub PII at the edge — never on your critical path.

Watch the self-healing loop

Violations appear in Activity within a second. Open one to see the explanation, prove the suggested fix against the failing case, then adopt the hardened prompt — confirmed against your live traffic.

The universal webhook

Any stack that can POST JSON works. Fire this after each agent turn:

curl -X POST https://sentinelproxy.tomphillips.uk/api/v1/ingest \
  -H "X-API-Key: <your project key>" \
  -H "Content-Type: application/json" \
  -d '{
    "system_prompt": "<your agent system prompt>",
    "user_input": "<the user message>",
    "llm_response": "<the model output>",
    "latency_ms": 800
  }'

How it works

The judge

Every exchange is graded by Claude against your guardrails. No rules to write — it reads intent, not regexes.

Guardrails

Four rule types — hallucination, prompt injection, PII leak, competitor mention — each toggleable per project with a severity.

Detect → Prove → Heal → Confirm

For each violation you get a copy-paste prompt fix, proven against the failing case and the whole failure class, then confirmed against production traffic before you adopt it.

Batch pattern-healing & PromptBreaker

Heal a whole class of failures in one pass, and red-team your prompt with the PromptBreaker attack battery before attackers do.

Integrations

Native plugins for every major agent platform — install the package or node, point it at your project key, and the loop starts.

n8n

Add the node after your LLM / AI Agent step.

Settings → Community Nodes → Install
n8n-nodes-sentinelproxy

LangChain · Python

Attach SentinelProxyCallbackHandler via callbacks.

pip install "sentinelproxy[langchain]"

LangChain · JS / TS

Pass SentinelProxyCallbackHandler in the callbacks array.

npm i sentinelproxy-langchain

LiteLLM

litellm.success_callback = [sentinel_litellm_callback('sk_…')]

pip install sentinelproxy

Dify

Add the Log Agent Turn node to any Chatflow or Workflow.

Dify Marketplace → SentinelProxy

Alerts

In a project's Connect tab, paste a Slack incoming webhook (or any JSON endpoint — n8n, Zapier, Make, your own service). The moment the judge flags a violation, we POST a message with the project, the violation type and a link straight to it. Saving sends a test alert so you know it works.

FAQ

Will this slow down my agent?

No. Ingestion is async — we acknowledge with a 200 immediately and run the judge off your critical path. Your agent never waits on us.

What happens to PII?

Emails, phone numbers and card numbers are scrubbed at the edge before anything is written to the database (UK GDPR / DPA 2018 data minimisation).

Which model does the judge use?

Claude Haiku 4.5 by default. On a paid plan you can bring your own Anthropic key and pick your own judge model (Haiku, Sonnet or Opus) — for 25% off and uncapped evaluations.

How are evaluations counted?

Each judged turn is one evaluation. Free includes 500/month; paid plans scale up. Bring your own key and evaluations are uncapped (they bill to your Anthropic account).

Is my API key safe?

Treat the project key like a secret — keep it server-side. Any BYOK Anthropic key is stored encrypted (AES-256-GCM), never returned to the browser, and decrypted only server-side.

Ready to wire it in?

Create a project, drop in one plugin, and watch the first violation heal itself.

Get started free