ramen ai

Latest ecosystem updates

Integrations

Governance that plugs into the stack you already run.

Every plugin attaches to a native seam in your framework, evaluates the decoded tool payload pre-execution, and halts the call on a blocked verdict. No proxy to operate, no model to retrain, no rewrite of your agent.

Browse the source Mint a free API key

Free Starter tier — 1,000 evaluations/month, BYOK, full Core IT Security bundle.

Official Plugins

Six integrations across agents, models, and datasets.

Each plugin verifies the Ed25519 receipt locally, so a blocked verdict is cryptographically provable at the point of enforcement — not just logged.

  • Python
    Agent Framework

    LangChain — view plugin source on GitHub (opens in a new tab)

    A callback handler that intercepts agent tool calls pre-execution, evaluates the serialized tool definition and input, and halts the chain on a blocked verdict.

    pip install -e plugins/langchain-python
    RamenSafetyCallbackHandler → raises RamenSafetyException
    Python ≥ 3.10 · langchain-core ≥ 0.2
  • Python
    Agent Framework

    PydanticAI — view plugin source on GitHub (opens in a new tab)

    Middleware that hooks PydanticAI's native args_validator, evaluating the tool name and resolved arguments before the tool body ever runs.

    pip install -e plugins/pydantic-ai
    ramen_firewall() → @agent.tool(args_validator=…)
    Python ≥ 3.10 · pydantic-ai ≥ 0.0.14
  • TypeScript
    Transport Proxy

    Model Context Protocol — view plugin source on GitHub (opens in a new tab)

    A universal MCP stdio proxy that intercepts tools/call JSON-RPC messages at the transport layer and blocks malicious payloads before they reach the downstream server.

    npx @ramen-ai/mcp-shield-proxy --help
    Drop-in stdio proxy — no SDK changes required
    Any MCP server using stdio transport
  • TypeScript
    Orchestration

    Microsoft AGT — view plugin source on GitHub (opens in a new tab)

    Governs any agent running under the Microsoft Agent Governance Toolkit, evaluating proposed tool calls and returning deterministic steering to the host agent on a block.

    npm install @ramen-ai/agt-middleware
    RamenFirewallBackend → GovernanceDenied
    @microsoft/agent-governance-sdk
  • Python
    Classical ML / XGBoost

    ramen-mlflow-guard — view plugin source on GitHub (opens in a new tab)

    Algorithmic governance for MLflow-served classical models, including XGBoost workflows. Evaluates feature values and available SHAP attributions for proxy bias before inference.

    pip install ramen-mlflow-guard
    RamenGovernedModel → GovernanceDeniedException
    Python ≥ 3.10 · MLflow ≥ 2.9
  • Python
    Dataset Distillation

    ramen-data-filter — view plugin source on GitHub (opens in a new tab)

    Dual-mode Pandas and CSV filtration for RAG ingestion and MLOps datasets, with strict exclusion or operator-controlled semantic imputation for blocked rows.

    pip install ramen-data-filter
    filter_dataframe() / filter_csv()
    Python ≥ 3.10 · Pandas 2.3.3

Client Libraries

Or build directly against the API.

Published to the public registries. Each client verifies the Ed25519 receipt locally, so you are not trusting our word for a verdict.

Beyond the Agent Runtime

Enforcement in the pipeline and at the attestation boundary.

GitHub Actions (opens in a new tab)

Scans pull request diffs for AI prompt modifications, evaluates the added text, and fails the build on a blocked verdict with a receipted PR comment.

cMCP + TRACE (opens in a new tab)

Officially merged upstream in the Linux Foundation's AgentTrust ecosystem, this cMCP policy adapter maps each verified V5 receipt onto a TRACE Trust Record for cross-vendor attestation.

Common Contract

Every plugin follows the same three-step contract.

1

Intercept

The plugin attaches to a native hook and captures the decoded tool name and arguments before the tool body executes.

2

Evaluate

The payload is scored for semantic intent against your policy bundles over POST /api/v1/paas/evaluate. Stateless — nothing is retained.

3

Halt or allow

A blocked verdict raises in-process and returns deterministic steering, so the agent replans instead of failing. The V5 receipt is verified locally.

Open Challenge

Can you bypass it?

Keyword filters catch syntax and fail on encoded payloads and corporate jargon. Our Red Team Guide publishes the zero-day evasion vectors we test against — including the Morse-code-encoded transfer used in the Grok/Bankr simulation. Every verdict in it was checked against the live API before publication.

Read the Red Team Guide (opens in a new tab)

Wire it up in an afternoon.

Mint a key on the free Starter tier, point a plugin at your policy bundle, and watch the first blocked tool call land with a signed receipt.

Get an API key Read the docs