Runnable, copy-pasteable recipes for putting Highflame security in front of your AI agents and LLM traffic — without rewriting your stack. Each recipe pairs a short setup you do in Highflame Studio with a script you can run to watch the guardrail work for yourself.
Looking for concepts, architecture, and product docs? Those live at docs.highflame.ai. This repo is the hands-on, run-it-yourself companion.
There's no single "correct" way to adopt Highflame — the right path depends on how your team already runs AI. Find your row:
| You run… | Recipe | What you change | What you gain |
|---|---|---|---|
| The Highflame SDK directly | recipes/sdk/ |
pip install highflame + four lines |
Full guardrail coverage on every prompt and tool call; the foundation all other recipes build on |
| Coding agents (Claude Code, Cursor, Codex…) behind Tailscale Aperture | recipes/aperture/ |
Add one Highflame hook in Aperture | Block secret & PII leaks, redact PII, stop prompt injection — with per-developer identity on every request |
| LiteLLM already | recipes/litellm/ |
Add Highflame as an upstream provider, or as a guardrail hook | Keep your routing and budgets; add the security + identity layer |
| AI agents from an IdP (Google Workspace, Okta, Entra…) | recipes/agent-governance/ |
Connect your IdP; adopt agents; attach a guardrail | Discover every agent, give each an accountable owner, and govern what it does at runtime |
| The OpenAI SDK / LangChain from scratch | coming soon | Point your base URL at Highflame | Policy enforcement + observability, zero instrumentation |
| Your own agents / framework | coming soon | A guard call + a per-agent identity | Inline guardrails and per-agent identity |
Each row is one self-contained recipe directory — read only the one you need.
Every recipe uses the same two things:
- A Highflame account — sign in at studio.highflame.ai.
- An API key — generate one in Studio; each recipe's README says exactly where.
Recipes read the key from an environment variable and ship a .env.example. Never commit
a real key.
| Recipe | For | Format | Status |
|---|---|---|---|
| Highflame SDK | Evaluate prompts & tools directly; the foundation | Marimo notebooks | ✅ ready |
| Code agents via Tailscale Aperture | Securing Claude Code / Cursor / Codex behind Aperture | Python scripts | ✅ ready |
| LiteLLM | Teams already running LiteLLM | Python scripts | ✅ ready |
| Overwatch policy catalog | What Overwatch catches for IDE coding agents (Cursor, Claude Code, Copilot) | Reference doc | ✅ ready |
| Agent governance | Discovering, adopting & guarding agents from Google Workspace / Okta / Entra | Python script | ✅ ready |
| OpenAI SDK / LangChain (greenfield) | New projects | — | coming soon |
| Portkey | Teams on Portkey | — | coming soon |
Every recipe follows the same shape, so a five-minute walkthrough looks the same each time:
- Set up once in Studio — generate a key and turn on a policy (each recipe has the exact click-path, with screenshots).
- Run the proof — a short script (or Marimo notebook) sends a representative request and shows you Highflame's decision: allow, deny (with the message you set in Studio), or modify (PII redacted).
cd recipes/<recipe>
cp .env.example .env # add your Highflame API key
pip install -r requirements.txt
python <script named in the recipe README>cd recipes/sdk
pip install -r requirements.txt
cp .env.example .env # add your HIGHFLAME_API_KEY
marimo run 01_quickstart.py # interactive browser UIOr open any notebook in the full editor:
marimo edit 01_quickstart.pyThe SDK recipe (recipes/sdk/) is a series of four interactive
Marimo notebooks — reactive Python cells you run and
modify in the browser.
| Notebook | What it covers |
|---|---|
01_quickstart.py |
Connect, evaluate a prompt, @shield.prompt decorator |
02_tool_security.py |
Guard tool calls, @shield.tool, rich ToolContext |
03_agentic_sessions.py |
Cross-turn session tracking, full OpenAI agent loop |
04_wave_d_decisions.py |
All five AARM decisions: allow / deny / modify / step_up / defer |
Contributing to this repo? See CONTRIBUTING.md.