yagami

Yagami

Open-source AI context firewall for governed model, retrieval, memory, and tool access.

Documentation | Gateway API | Deployment | Security | Roadmap

Yagami sits between your software and local models, cloud LLMs, retrieval systems, and tools. It classifies context locally, evaluates versioned policy, routes only to allowed destinations, inspects outputs, and produces content-free evidence for each decision.

Existing OpenAI SDK applications can adopt it by changing one base_url. Yagami can run as a headless gateway, in a container or Kubernetes, or with its included React control surface.

Try it in 60 seconds

The demo requires no API key, provider account, Ollama model, or Node.js:

python -m pip install yagami
yagami demo

Open http://127.0.0.1:8000. Demo mode uses a local echo backend, blocks cloud routing, and exercises the UI, policy, lineage, storage, and audit path.

Watch the two-minute demo.

Take the no-data security tour or run the flagship security demos for secret containment, poisoned retrieval, and identity-bound tool approval.

Protect an application

Initialize persistent user configuration, check the host, and start Yagami:

yagami init
yagami doctor
yagami serve

Then point an OpenAI client at the gateway:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="your-yagami-project-key",
)

response = client.chat.completions.create(
    model="yagami-auto",
    messages=[{"role": "user", "content": "Summarize this document."}],
    metadata={
        "purpose": "internal-documentation",
        "sensitivity": "none",
        "session_id": "example-session",
    },
)
print(response.choices[0].message.content)

Supported caller sensitivity values are none, phi, phi_medical, and secret. A caller hint can make the policy stricter; it cannot lower a sensitivity detected by Yagami.

For production authentication, policy, and deployment settings, follow the deployment guide.

Why teams use Yagami

• Deterministic containment after classification. Once context is labeled as PHI or secret, default policy permits local backends only. Sensitive history and tool results inherit the same restriction.
• One governed data plane. Chat Completions, Responses, the browser chat, and MCP use the same policy, lineage, transformation, output-DLP, budget, and audit pipeline.
• Policy as code. Preview and replay decisions, run regression cases in CI, and promote deterministic Ed25519-signed policy bundles.
• Evidence without prompt logging. Policy passports, hash-chained audit records, Prometheus metrics, and OpenTelemetry spans carry labels, hashes, IDs, and counts rather than prompt or completion content.
• Model choice without policy duplication. Route to local engines, direct cloud providers, or an existing OpenAI-compatible gateway behind one enforcement point.
• Governed tools. Evaluate function tools and MCP calls before execution, require short-lived one-time approvals, and keep inbound credentials from being forwarded to downstream servers.

Core capabilities

Area	Included
Compatible APIs	OpenAI Chat Completions, core Responses API, Streamable HTTP MCP
Identity	Scoped project API keys and OIDC/JWT workload identity
Policy	Versioned YAML/JSON rules, restrictive merging, preview, replay, shadow mode, regression tests, signed bundles
Privacy	Local classification, caller sensitivity, context lineage, AES-GCM tokenization, rehydration, output DLP, optional Presidio
Tools	Function calling, governed built-in skills, stdio and remote MCP, one-time approvals
Operations	Spend/rate/concurrency/context limits, health checks, Prometheus, OpenTelemetry, SIEM export, approval webhooks
Packaging	Python 3.11-3.14, PyPI, non-root container, Docker Compose, Helm, SBOMs, checksums, and build provenance

Models and integrations

Local generation works with Ollama, llama.cpp through the optional llama-cpp-python runtime, and Microsoft Foundry Local through its loopback OpenAI-compatible service. Direct cloud adapters cover Anthropic, OpenAI, Mistral, Groq, OpenRouter, Google Gemini, and Stability AI image generation.

Yagami also works with LangChain/LangGraph, the Vercel AI SDK, Microsoft Presidio, Splunk HEC and generic SIEM webhooks, Slack and Teams approval notifications, and upstream gateways such as LiteLLM, Portkey, Kong, or Envoy. See the integration recipes.

How enforcement works

application or agent
  -> authentication and project limits
  -> local sensitivity and context-lineage inspection
  -> versioned policy and optional transformation
  -> allowed local model, cloud model, retrieval source, or tool
  -> output DLP
  -> response plus content-free policy passport and audit evidence

Policy is the final authority. Slash commands and explicit backend selection cannot override a sensitive-data restriction. Classifier failures fail local by default, and cloud routes can be blocked entirely or stopped at a daily spend cap.

Important limitations

Yagami is an enforcement component, not a compliance certification. Automated detection can miss sensitive data. Strict deployments should declare sensitivity at the caller, use a local-only policy, test organization-specific cases, encrypt storage at the host or volume layer, and review the threat model.

The project is alpha. Validate policy and failure behavior against your own requirements before production use.

Documentation

• Start here
• Gateway API
• Integrations
• Deployment
• Security and threat model

Contributing

Focused issues and pull requests are welcome. Read CONTRIBUTING.md, the security policy, and the code of conduct.

License

MIT - Copyright (c) 2026 Matthew Tracy and Yagami contributors.