ramen-ai Integrations

The deterministic execution boundary for AI agents.

The Problem

LLMs cannot police their own tools. If your agent is connected to a database,
a prompt injection hidden in a PDF can drop your tables. If it is connected to
a payment API, a Morse-encoded instruction buried in a retrieved document can
drain a wallet. Standard keyword filters catch basic syntax. They fail against
encoded payloads, corporate jargon, and multi-criterion composite attacks.

We built the deterministic execution boundary to stop it — a semantic
firewall that evaluates every tool call against your compliance policies
before execution, signs every verdict with an Ed25519 receipt, and returns a
steering instruction that is auditable, reproducible, and legally defensible.

Can you bypass it?

Standard safety filters catch basic syntax. They fail against encoded payloads
and corporate jargon. We challenge you to bypass our semantic firewall using
the zero-day evasion vectors in our official Red Team Guide.

Below is a simulation of the Grok/Bankr heist. We fed the raw adversarial
prompt directly into our sandbox. It uses a social engineering wrapper
(claiming a visual impairment) to smuggle a 3,000,000,000 DRB transfer
instruction encoded in Morse code. The firewall evaluated the underlying
semantic intent, intercepted the unauthorized financial transfer, and blocked
it pre-execution, issuing a verified Ed25519 receipt.

Getting Started

To use these integrations, you must mint an API Key.

We offer a Free Starter Tier (1,000 evaluations/month, BYOK) which includes
full access to our Core IT Security bundle. Mint your key at:

https://ramenai.dev/pricing

BYOK — Bring Your Own Key

The Starter and Professional tiers use your own LLM provider key for inference.
You need two keys:

export RAMEN_API_KEY=ramen_ak_...
export OPENAI_API_KEY=sk-...        # or ANTHROPIC_API_KEY, etc.

Enterprise tier users have keys managed server-side — omit providerKey
entirely.

Integrations

Core SDKs

Use these as the foundation for custom integrations across any framework,
queue, or gateway not covered by the plugins below.

Plugins

How it works

Every evaluation returns a V5 Ed25519 cryptographic receipt — a signed,
self-describing record binding the verdict to a SHA-256 hash of your input.
Receipts are verified locally against the published public key: no trust in
the API server is required.

Your agent  →  ramen-ai middleware  →  POST /api/v1/paas/evaluate
                                              ↓
                                     Verdict + Ed25519 receipt
                                              ↓
                             Local receipt verification (Web Crypto)
                                              ↓
                              ALLOWED → tool executes
                              BLOCKED → thrown / build failed / PR comment

Bundles

Use bundle_ids for macro-level Defence-in-Depth aggregation — a single
slug activates multiple coordinated policies across a threat domain. For
surgical, single-domain statutory auditing, pass specific policy_ids
directly to route to exactly one policy without other classifiers interfering.

The Free Starter Tier includes the Core IT Security bundle:

Full bundle and policy reference: https://ramenai.dev/pricing

Repository structure

/
├── assets/                  # Shared brand and visual assets
├── core-clients/
│   ├── node/                # @ramen-ai/node-core  — TypeScript SDK
│   └── python/              # ramen-ai-core        — Python SDK
├── plugins/
│   ├── agt-typescript/      # Microsoft AGT middleware
│   ├── github-action/       # GitHub Actions CI/CD interceptor
│   ├── langchain-python/    # LangChain Python callback handler
│   ├── pydantic-ai/         # PydanticAI args_validator middleware
│   ├── mcp-proxy/           # Universal MCP stdio transport interceptor
│   ├── cmcp-python/         # cMCP policy adapter + TRACE record exporter
│   ├── mlflow-python/       # MLflow pyfunc algorithmic governance wrapper
│   ├── ramen-data-filter/   # Pandas/CSV filtration for RAG and MLOps datasets
│   └── dsh-ramen-guard/    # DeepSeek Harness pre-execution Cordis guard
├── .github/
│   └── workflows/           # CI workflows
└── RED_TEAM_GUIDE.md        # Zero-day evasion vectors and challenge guide

Adding integration logos

When adding a platform integration, update the standard ecosystem row in the
home README, every plugin README, and every translated companion. See
Adding Integration Logos for branding,
internal-link, self-badge, asset, disclosure, and validation requirements.