halyn 2.1.1__tar.gz → 2.1.3__tar.gz

{halyn-2.1.1/src/halyn.egg-info → halyn-2.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: halyn
-Version: 2.1.1
+Version: 2.1.3
 Summary: Halyn — The governance layer for AI agents. Every action intercepted. Every decision auditable.
 Author-email: Elmadani SALKA <contact@halyn.dev>
 License: BSL-1.1
@@ -157,18 +157,84 @@ Deployed via `/etc/opt/chrome/policies/managed/halyn.json` — the agent cannot
 
 ---
 
-## Supported Agents
-
-| Agent | Type | Integration |
-|-------|------|-------------|
-| Claude Sonnet 4.6 / Opus 4.6 (Anthropic) | Cloud | Proxy + MCP |
-| GPT-4.1 / o3 (OpenAI) | Cloud | Proxy |
-| Gemini 3.1 Pro / Flash (Google) | Cloud | Proxy |
-| Ollama | Local | Direct |
-| LM Studio | Local | Direct |
-| Jan.ai | Local | Direct |
-| OpenClaw | Agentic | Interceptor |
-| Any MCP agent | Any | MCP server |
+## Compatible AI
+
+Halyn intercepts at the kernel and proxy level. It does not care which AI is running — it audits all of them equally. No AI is excluded.
+
+### How compatibility works
+
+Halyn intercepts three things:
+- **API calls** (iptables REDIRECT on port 443/80) — catches any HTTP request to any AI provider
+- **Filesystem access** (inotify/FSEvents/eBPF) — catches any agent touching files, regardless of origin
+- **Process syscalls** (eBPF, Linux ≥5.8) — catches any agent at the kernel level
+
+This means: if an AI agent makes an API call or accesses your system, Halyn sees it.
+
+### Cloud AI
+
+| Provider | Models (March 2026) | API |
+|----------|---------------------|-----|
+| **Anthropic** | Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 | api.anthropic.com |
+| **OpenAI** | GPT-4.1, GPT-4.1 mini, GPT-4.1 nano, o3, o4-mini | api.openai.com |
+| **Google** | Gemini 3.1 Pro, Gemini 3.1 Flash, Gemini 3.1 Flash-Lite | generativelanguage.googleapis.com |
+| **Mistral AI** | Mistral Large 2, Mistral Small 3, Codestral | api.mistral.ai |
+| **xAI** | Grok-3, Grok-3 mini | api.x.ai |
+| **DeepSeek** | DeepSeek-V3, DeepSeek-R1 | api.deepseek.com |
+| **Cohere** | Command R+, Command R, Aya | api.cohere.com |
+| **Perplexity** | Sonar Pro, Sonar, Sonar Reasoning | api.perplexity.ai |
+| **01.AI** | Yi-Large, Yi-Vision | api.01.ai |
+| **Alibaba** | Qwen-Max, Qwen-Plus, Qwen-Turbo | dashscope.aliyuncs.com |
+| **Baidu** | ERNIE 4.5, ERNIE Speed | aip.baidubce.com |
+| **Amazon Bedrock** | Claude, Titan, Llama, Mistral (via AWS) | bedrock.amazonaws.com |
+| **Azure OpenAI** | GPT-4.1, o3 (via Microsoft) | *.openai.azure.com |
+| **NVIDIA NIM** | Llama 3.3, Mistral, DeepSeek-R1 (on NVIDIA cloud) | integrate.api.nvidia.com |
+| **Together AI** | 50+ open models via API | api.together.xyz |
+| **Groq** | Llama, Mixtral, Gemma (ultra-fast inference) | api.groq.com |
+| **Fireworks AI** | Llama, Mixtral, DeepSeek | api.fireworks.ai |
+
+### Local AI
+
+Any local model is compatible — Halyn intercepts at the process level, not the network level.
+
+| Runtime | Models | Notes |
+|---------|--------|-------|
+| **Ollama** | Llama 3.3, Qwen2.5, Mistral, DeepSeek-R1, Phi-4, Gemma 3, ... | OpenAI-compatible API |
+| **LM Studio** | Any GGUF model | OpenAI-compatible server |
+| **Jan.ai** | Any GGUF or ONNX model | Desktop + server mode |
+| **GPT4All** | Llama, Mistral, Phi variants | Offline, no telemetry |
+| **llama.cpp** | Any GGUF model directly | Server mode (`--server`) |
+| **LocalAI** | 100+ models, any GGUF | Drop-in OpenAI replacement |
+| **text-generation-webui** | Any HuggingFace model | Extension ecosystem |
+| **KoboldCpp** | Any GGUF model | Focus on creative writing |
+| **OpenWebUI** | Ollama + OpenAI frontend | Browser-based |
+| **AnythingLLM** | Multi-model workspace | Team-friendly |
+| **Xinference** | HuggingFace + GGUF | Enterprise local inference |
+| **vLLM** | HuggingFace models | High-throughput server |
+| **TGI (HuggingFace)** | HuggingFace models | Production inference |
+
+### Agentic frameworks
+
+Halyn intercepts any agentic system. The agent framework doesn't matter.
+
+| Framework | Notes |
+|-----------|-------|
+| **OpenClaw** | Full interceptor — every action audited |
+| **Claude Cowork** | Proxy + filesystem hooks |
+| **Claude Code** | Process-level monitoring |
+| **LangChain** | API calls intercepted automatically |
+| **LlamaIndex** | API calls intercepted automatically |
+| **AutoGen** | API calls intercepted automatically |
+| **CrewAI** | API calls intercepted automatically |
+| **Semantic Kernel** | API calls intercepted automatically |
+| **BeeQ** | Native AAP integration |
+| **Any MCP agent** | MCP server passthrough |
+| **Any A2A agent** | Network-level interception |
+| **Any OpenAI-compatible API** | Universal proxy compatibility |
+
+### The rule
+
+> If an AI touches your machine or calls an API — Halyn sees it.
+> No exception. No exclusion. That's the point.
 
 ---
 

{halyn-2.1.1 → halyn-2.1.3}/README.md

@@ -124,18 +124,84 @@ Deployed via `/etc/opt/chrome/policies/managed/halyn.json` — the agent cannot
 
 ---
 
-## Supported Agents
-
-| Agent | Type | Integration |
-|-------|------|-------------|
-| Claude Sonnet 4.6 / Opus 4.6 (Anthropic) | Cloud | Proxy + MCP |
-| GPT-4.1 / o3 (OpenAI) | Cloud | Proxy |
-| Gemini 3.1 Pro / Flash (Google) | Cloud | Proxy |
-| Ollama | Local | Direct |
-| LM Studio | Local | Direct |
-| Jan.ai | Local | Direct |
-| OpenClaw | Agentic | Interceptor |
-| Any MCP agent | Any | MCP server |
+## Compatible AI
+
+Halyn intercepts at the kernel and proxy level. It does not care which AI is running — it audits all of them equally. No AI is excluded.
+
+### How compatibility works
+
+Halyn intercepts three things:
+- **API calls** (iptables REDIRECT on port 443/80) — catches any HTTP request to any AI provider
+- **Filesystem access** (inotify/FSEvents/eBPF) — catches any agent touching files, regardless of origin
+- **Process syscalls** (eBPF, Linux ≥5.8) — catches any agent at the kernel level
+
+This means: if an AI agent makes an API call or accesses your system, Halyn sees it.
+
+### Cloud AI
+
+| Provider | Models (March 2026) | API |
+|----------|---------------------|-----|
+| **Anthropic** | Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 | api.anthropic.com |
+| **OpenAI** | GPT-4.1, GPT-4.1 mini, GPT-4.1 nano, o3, o4-mini | api.openai.com |
+| **Google** | Gemini 3.1 Pro, Gemini 3.1 Flash, Gemini 3.1 Flash-Lite | generativelanguage.googleapis.com |
+| **Mistral AI** | Mistral Large 2, Mistral Small 3, Codestral | api.mistral.ai |
+| **xAI** | Grok-3, Grok-3 mini | api.x.ai |
+| **DeepSeek** | DeepSeek-V3, DeepSeek-R1 | api.deepseek.com |
+| **Cohere** | Command R+, Command R, Aya | api.cohere.com |
+| **Perplexity** | Sonar Pro, Sonar, Sonar Reasoning | api.perplexity.ai |
+| **01.AI** | Yi-Large, Yi-Vision | api.01.ai |
+| **Alibaba** | Qwen-Max, Qwen-Plus, Qwen-Turbo | dashscope.aliyuncs.com |
+| **Baidu** | ERNIE 4.5, ERNIE Speed | aip.baidubce.com |
+| **Amazon Bedrock** | Claude, Titan, Llama, Mistral (via AWS) | bedrock.amazonaws.com |
+| **Azure OpenAI** | GPT-4.1, o3 (via Microsoft) | *.openai.azure.com |
+| **NVIDIA NIM** | Llama 3.3, Mistral, DeepSeek-R1 (on NVIDIA cloud) | integrate.api.nvidia.com |
+| **Together AI** | 50+ open models via API | api.together.xyz |
+| **Groq** | Llama, Mixtral, Gemma (ultra-fast inference) | api.groq.com |
+| **Fireworks AI** | Llama, Mixtral, DeepSeek | api.fireworks.ai |
+
+### Local AI
+
+Any local model is compatible — Halyn intercepts at the process level, not the network level.
+
+| Runtime | Models | Notes |
+|---------|--------|-------|
+| **Ollama** | Llama 3.3, Qwen2.5, Mistral, DeepSeek-R1, Phi-4, Gemma 3, ... | OpenAI-compatible API |
+| **LM Studio** | Any GGUF model | OpenAI-compatible server |
+| **Jan.ai** | Any GGUF or ONNX model | Desktop + server mode |
+| **GPT4All** | Llama, Mistral, Phi variants | Offline, no telemetry |
+| **llama.cpp** | Any GGUF model directly | Server mode (`--server`) |
+| **LocalAI** | 100+ models, any GGUF | Drop-in OpenAI replacement |
+| **text-generation-webui** | Any HuggingFace model | Extension ecosystem |
+| **KoboldCpp** | Any GGUF model | Focus on creative writing |
+| **OpenWebUI** | Ollama + OpenAI frontend | Browser-based |
+| **AnythingLLM** | Multi-model workspace | Team-friendly |
+| **Xinference** | HuggingFace + GGUF | Enterprise local inference |
+| **vLLM** | HuggingFace models | High-throughput server |
+| **TGI (HuggingFace)** | HuggingFace models | Production inference |
+
+### Agentic frameworks
+
+Halyn intercepts any agentic system. The agent framework doesn't matter.
+
+| Framework | Notes |
+|-----------|-------|
+| **OpenClaw** | Full interceptor — every action audited |
+| **Claude Cowork** | Proxy + filesystem hooks |
+| **Claude Code** | Process-level monitoring |
+| **LangChain** | API calls intercepted automatically |
+| **LlamaIndex** | API calls intercepted automatically |
+| **AutoGen** | API calls intercepted automatically |
+| **CrewAI** | API calls intercepted automatically |
+| **Semantic Kernel** | API calls intercepted automatically |
+| **BeeQ** | Native AAP integration |
+| **Any MCP agent** | MCP server passthrough |
+| **Any A2A agent** | Network-level interception |
+| **Any OpenAI-compatible API** | Universal proxy compatibility |
+
+### The rule
+
+> If an AI touches your machine or calls an API — Halyn sees it.
+> No exception. No exclusion. That's the point.
 
 ---
 

{halyn-2.1.1 → halyn-2.1.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "halyn"
-version = "2.1.1"
+version = "2.1.3"
 description = "Halyn — The governance layer for AI agents. Every action intercepted. Every decision auditable."
 requires-python = ">=3.10"
 license = {text = "BSL-1.1"}

{halyn-2.1.1 → halyn-2.1.3}/src/halyn/__init__.py

@@ -9,7 +9,7 @@ Every action intercepted. Every decision auditable.
 The AI cannot bypass it.
 """
 
-__version__ = "2.1.1"
+__version__ = "2.1.3"
 __author__ = "Elmadani SALKA"
 __license__ = "BSL-1.1"
 __email__ = "contact@halyn.dev"

halyn-2.1.3/src/halyn/llm.py

@@ -0,0 +1,362 @@
+# Copyright (c) 2026 Elmadani SALKA
+# Licensed under BSL-1.1. See LICENSE file.
+# Commercial use requires a license — contact@halyn.dev
+
+"""
+LLM Connector — Multi-provider abstraction.
+
+Halyn is provider-agnostic. It monitors any AI, regardless of origin.
+This module provides connectors for direct LLM integration (optional).
+The proxy layer works independently — without any connector configured.
+
+Supported by the proxy (no connector needed):
+  Cloud: Anthropic, OpenAI, Google, Mistral, xAI, DeepSeek, Cohere,
+         Perplexity, 01.AI, Alibaba, Baidu, Amazon Bedrock, Azure OpenAI,
+         NVIDIA NIM, Together AI, Groq, Fireworks AI, and any HTTP API.
+  Local: Ollama, LM Studio, Jan.ai, GPT4All, llama.cpp, LocalAI,
+         text-generation-webui, KoboldCpp, vLLM, TGI, and any local server.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import urllib.request
+from abc import ABC, abstractmethod
+from typing import Any
+
+log = logging.getLogger(__name__)
+
+
+# ─── Base ─────────────────────────────────────────────────────────────────────
+
+class LLMResponse:
+    def __init__(self, content: str, model: str = "", usage: dict | None = None) -> None:
+        self.content = content
+        self.model = model
+        self.usage = usage or {}
+
+
+class LLMConnector(ABC):
+    """Base class for all LLM connectors."""
+
+    @abstractmethod
+    def complete(
+        self,
+        messages: list[dict],
+        system: str = "",
+        max_tokens: int = 1024,
+    ) -> LLMResponse:
+        ...
+
+
+# ─── Cloud providers ──────────────────────────────────────────────────────────
+
+class AnthropicConnector(LLMConnector):
+    """Anthropic — Claude Sonnet 4.6, Opus 4.6, Haiku 4.5."""
+
+    def __init__(self, api_key: str = "", model: str = "claude-sonnet-4-6") -> None:
+        self.api_key = api_key or os.environ.get("ANTHROPIC_API_KEY", "")
+        self.model = model
+        self.endpoint = "https://api.anthropic.com/v1/messages"
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        body: dict[str, Any] = {"model": self.model, "max_tokens": max_tokens, "messages": messages}
+        if system:
+            body["system"] = system
+        data = self._post(self.endpoint, body, {
+            "x-api-key": self.api_key,
+            "anthropic-version": "2023-06-01",
+        })
+        content = data["content"][0]["text"]
+        return LLMResponse(content, self.model, data.get("usage"))
+
+    def _post(self, url: str, body: dict, headers: dict) -> dict:
+        payload = json.dumps(body).encode()
+        req = urllib.request.Request(url, data=payload, method="POST")
+        req.add_header("Content-Type", "application/json")
+        for k, v in headers.items():
+            req.add_header(k, v)
+        with urllib.request.urlopen(req, timeout=60) as resp:
+            return json.loads(resp.read())
+
+
+class OpenAIConnector(LLMConnector):
+    """OpenAI — GPT-4.1, GPT-4.1 mini, GPT-4.1 nano, o3, o4-mini.
+    Also compatible with: Azure OpenAI, NVIDIA NIM, Together AI, Groq,
+    Fireworks AI, and any OpenAI-compatible endpoint.
+    """
+
+    def __init__(
+        self,
+        api_key: str = "",
+        model: str = "gpt-4.1",
+        endpoint: str = "https://api.openai.com/v1",
+    ) -> None:
+        self.api_key = api_key or os.environ.get("OPENAI_API_KEY", "")
+        self.model = model
+        self.endpoint = endpoint.rstrip("/")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        msgs = ([{"role": "system", "content": system}] if system else []) + messages
+        body = {"model": self.model, "max_tokens": max_tokens, "messages": msgs}
+        data = self._post(f"{self.endpoint}/chat/completions", body)
+        content = data["choices"][0]["message"]["content"]
+        return LLMResponse(content, self.model, data.get("usage"))
+
+    def _post(self, url: str, body: dict) -> dict:
+        payload = json.dumps(body).encode()
+        req = urllib.request.Request(url, data=payload, method="POST")
+        req.add_header("Content-Type", "application/json")
+        req.add_header("Authorization", f"Bearer {self.api_key}")
+        with urllib.request.urlopen(req, timeout=60) as resp:
+            return json.loads(resp.read())
+
+
+class GeminiConnector(LLMConnector):
+    """Google — Gemini 3.1 Pro, Gemini 3.1 Flash, Gemini 3.1 Flash-Lite."""
+
+    def __init__(self, api_key: str = "", model: str = "gemini-3.1-flash-lite-preview") -> None:
+        self.api_key = api_key or os.environ.get("GOOGLE_API_KEY", "")
+        self.model = model
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        url = (
+            f"https://generativelanguage.googleapis.com/v1beta/models/"
+            f"{self.model}:generateContent?key={self.api_key}"
+        )
+        contents = [{"role": m["role"].replace("assistant", "model"), "parts": [{"text": m["content"]}]}
+                    for m in messages]
+        body: dict[str, Any] = {
+            "contents": contents,
+            "generationConfig": {"maxOutputTokens": max_tokens},
+        }
+        if system:
+            body["systemInstruction"] = {"parts": [{"text": system}]}
+        payload = json.dumps(body).encode()
+        req = urllib.request.Request(url, data=payload, method="POST")
+        req.add_header("Content-Type", "application/json")
+        with urllib.request.urlopen(req, timeout=60) as resp:
+            data = json.loads(resp.read())
+        content = data["candidates"][0]["content"]["parts"][0]["text"]
+        return LLMResponse(content, self.model)
+
+
+class MistralConnector(LLMConnector):
+    """Mistral AI — Mistral Large 2, Mistral Small 3, Codestral."""
+
+    def __init__(self, api_key: str = "", model: str = "mistral-small-latest") -> None:
+        self.api_key = api_key or os.environ.get("MISTRAL_API_KEY", "")
+        self.model = model
+        self._oa = OpenAIConnector(api_key, model, "https://api.mistral.ai/v1")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._oa.api_key = self.api_key
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class XAIConnector(LLMConnector):
+    """xAI — Grok-3, Grok-3 mini."""
+
+    def __init__(self, api_key: str = "", model: str = "grok-3-mini") -> None:
+        self.api_key = api_key or os.environ.get("XAI_API_KEY", "")
+        self._oa = OpenAIConnector(api_key, model, "https://api.x.ai/v1")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._oa.api_key = self.api_key
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class DeepSeekConnector(LLMConnector):
+    """DeepSeek — DeepSeek-V3, DeepSeek-R1."""
+
+    def __init__(self, api_key: str = "", model: str = "deepseek-chat") -> None:
+        self.api_key = api_key or os.environ.get("DEEPSEEK_API_KEY", "")
+        self._oa = OpenAIConnector(api_key, model, "https://api.deepseek.com/v1")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._oa.api_key = self.api_key
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class GroqConnector(LLMConnector):
+    """Groq — Llama, Mixtral, Gemma (ultra-fast inference)."""
+
+    def __init__(self, api_key: str = "", model: str = "llama-3.3-70b-versatile") -> None:
+        self.api_key = api_key or os.environ.get("GROQ_API_KEY", "")
+        self._oa = OpenAIConnector(api_key, model, "https://api.groq.com/openai/v1")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._oa.api_key = self.api_key
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class PerplexityConnector(LLMConnector):
+    """Perplexity — Sonar Pro, Sonar, Sonar Reasoning."""
+
+    def __init__(self, api_key: str = "", model: str = "sonar") -> None:
+        self.api_key = api_key or os.environ.get("PERPLEXITY_API_KEY", "")
+        self._oa = OpenAIConnector(api_key, model, "https://api.perplexity.ai")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._oa.api_key = self.api_key
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class CohereConnector(LLMConnector):
+    """Cohere — Command R+, Command R, Aya."""
+
+    def __init__(self, api_key: str = "", model: str = "command-r-plus") -> None:
+        self.api_key = api_key or os.environ.get("COHERE_API_KEY", "")
+        self.model = model
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        url = "https://api.cohere.com/v2/chat"
+        chat_history = [{"role": m["role"], "content": m["content"]} for m in messages[:-1]]
+        body: dict[str, Any] = {
+            "model": self.model,
+            "messages": [{"role": m["role"], "content": m["content"]} for m in messages],
+            "max_tokens": max_tokens,
+        }
+        if system:
+            body["system"] = system
+        payload = json.dumps(body).encode()
+        req = urllib.request.Request(url, data=payload, method="POST")
+        req.add_header("Content-Type", "application/json")
+        req.add_header("Authorization", f"Bearer {self.api_key}")
+        with urllib.request.urlopen(req, timeout=60) as resp:
+            data = json.loads(resp.read())
+        content = data["message"]["content"][0]["text"]
+        return LLMResponse(content, self.model)
+
+
+# ─── Local providers ──────────────────────────────────────────────────────────
+
+class OllamaConnector(LLMConnector):
+    """Ollama — any local model: Llama 3.3, Qwen2.5, Mistral, DeepSeek-R1,
+    Phi-4, Gemma 3, and hundreds more. OpenAI-compatible API.
+    """
+
+    def __init__(self, model: str = "llama3.2", host: str = "http://localhost:11434") -> None:
+        self.model = model
+        self._oa = OpenAIConnector("ollama", model, f"{host}/v1")
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class LocalAIConnector(LLMConnector):
+    """LocalAI / LM Studio / Jan.ai / KoboldCpp / text-generation-webui /
+    llama.cpp / vLLM / TGI — any OpenAI-compatible local server.
+    """
+
+    def __init__(
+        self,
+        model: str = "local-model",
+        host: str = "http://localhost:8080",
+        api_key: str = "local",
+    ) -> None:
+        self.model = model
+        self._oa = OpenAIConnector(api_key, model, host)
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        return self._oa.complete(messages, system, max_tokens)
+
+
+class HuggingFaceConnector(LLMConnector):
+    """HuggingFace — any model via transformers pipeline (fully local)."""
+
+    def __init__(self, model: str = "microsoft/Phi-4") -> None:
+        self.model_name = model
+        self._pipeline: Any = None
+
+    def _load(self) -> None:
+        if self._pipeline is None:
+            try:
+                from transformers import pipeline  # type: ignore
+            except ImportError as e:
+                raise ImportError("pip install transformers torch") from e
+            log.info("llm.loading model=%s", self.model_name)
+            self._pipeline = pipeline("text-generation", model=self.model_name, device_map="auto")
+            log.info("llm.loaded model=%s", self.model_name)
+
+    def complete(self, messages: list[dict], system: str = "", max_tokens: int = 1024) -> LLMResponse:
+        self._load()
+        prompt = "\n".join(f"{m['role']}: {m['content']}" for m in messages)
+        if system:
+            prompt = f"system: {system}\n{prompt}"
+        result = self._pipeline(prompt, max_new_tokens=max_tokens, do_sample=False)
+        text = result[0]["generated_text"][len(prompt):].strip()
+        return LLMResponse(text, self.model_name)
+
+
+# ─── Factory ──────────────────────────────────────────────────────────────────
+
+def create_connector(provider: str, **kwargs: Any) -> LLMConnector:
+    """
+    Create a connector by provider name.
+
+    Cloud providers: anthropic, openai, azure, google, gemini, mistral,
+                     xai, grok, deepseek, cohere, perplexity, groq,
+                     fireworks, together, nvidia, bedrock
+    Local providers: ollama, lmstudio, jan, localai, llamacpp, gpt4all,
+                     kobold, vllm, tgi, huggingface, openai-compatible
+    """
+    # Normalize aliases
+    aliases = {
+        # Anthropic
+        "anthropic": "anthropic", "claude": "anthropic",
+        # OpenAI
+        "openai": "openai", "gpt": "openai", "azure": "openai",
+        "nvidia": "openai", "together": "openai", "fireworks": "openai",
+        "bedrock": "openai",  # via OpenAI-compat gateway
+        # Google
+        "google": "google", "gemini": "google",
+        # Mistral
+        "mistral": "mistral",
+        # xAI
+        "xai": "xai", "grok": "xai",
+        # DeepSeek
+        "deepseek": "deepseek",
+        # Groq
+        "groq": "groq",
+        # Perplexity
+        "perplexity": "perplexity", "sonar": "perplexity",
+        # Cohere
+        "cohere": "cohere", "command": "cohere",
+        # Local
+        "ollama": "ollama",
+        "lmstudio": "local", "jan": "local", "janai": "local",
+        "localai": "local", "llamacpp": "local", "llama.cpp": "local",
+        "kobold": "local", "koboldcpp": "local",
+        "vllm": "local", "tgi": "local",
+        "gpt4all": "local", "openwebui": "local",
+        "local": "local", "openai-compatible": "local",
+        # HuggingFace
+        "huggingface": "hf", "hf": "hf", "transformers": "hf",
+    }
+
+    connectors = {
+        "anthropic": AnthropicConnector,
+        "openai": OpenAIConnector,
+        "google": GeminiConnector,
+        "mistral": MistralConnector,
+        "xai": XAIConnector,
+        "deepseek": DeepSeekConnector,
+        "groq": GroqConnector,
+        "perplexity": PerplexityConnector,
+        "cohere": CohereConnector,
+        "ollama": OllamaConnector,
+        "local": LocalAIConnector,
+        "hf": HuggingFaceConnector,
+    }
+
+    key = aliases.get(provider.lower(), provider.lower())
+    cls = connectors.get(key)
+    if cls is None:
+        raise ValueError(
+            f"Unknown provider '{provider}'. "
+            f"Note: the Halyn proxy works with ANY AI regardless of this connector."
+        )
+    return cls(**kwargs)

{halyn-2.1.1 → halyn-2.1.3/src/halyn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: halyn
-Version: 2.1.1
+Version: 2.1.3
 Summary: Halyn — The governance layer for AI agents. Every action intercepted. Every decision auditable.
 Author-email: Elmadani SALKA <contact@halyn.dev>
 License: BSL-1.1
@@ -157,18 +157,84 @@ Deployed via `/etc/opt/chrome/policies/managed/halyn.json` — the agent cannot
 
 ---
 
-## Supported Agents
-
-| Agent | Type | Integration |
-|-------|------|-------------|
-| Claude Sonnet 4.6 / Opus 4.6 (Anthropic) | Cloud | Proxy + MCP |
-| GPT-4.1 / o3 (OpenAI) | Cloud | Proxy |
-| Gemini 3.1 Pro / Flash (Google) | Cloud | Proxy |
-| Ollama | Local | Direct |
-| LM Studio | Local | Direct |
-| Jan.ai | Local | Direct |
-| OpenClaw | Agentic | Interceptor |
-| Any MCP agent | Any | MCP server |
+## Compatible AI
+
+Halyn intercepts at the kernel and proxy level. It does not care which AI is running — it audits all of them equally. No AI is excluded.
+
+### How compatibility works
+
+Halyn intercepts three things:
+- **API calls** (iptables REDIRECT on port 443/80) — catches any HTTP request to any AI provider
+- **Filesystem access** (inotify/FSEvents/eBPF) — catches any agent touching files, regardless of origin
+- **Process syscalls** (eBPF, Linux ≥5.8) — catches any agent at the kernel level
+
+This means: if an AI agent makes an API call or accesses your system, Halyn sees it.
+
+### Cloud AI
+
+| Provider | Models (March 2026) | API |
+|----------|---------------------|-----|
+| **Anthropic** | Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 | api.anthropic.com |
+| **OpenAI** | GPT-4.1, GPT-4.1 mini, GPT-4.1 nano, o3, o4-mini | api.openai.com |
+| **Google** | Gemini 3.1 Pro, Gemini 3.1 Flash, Gemini 3.1 Flash-Lite | generativelanguage.googleapis.com |
+| **Mistral AI** | Mistral Large 2, Mistral Small 3, Codestral | api.mistral.ai |
+| **xAI** | Grok-3, Grok-3 mini | api.x.ai |
+| **DeepSeek** | DeepSeek-V3, DeepSeek-R1 | api.deepseek.com |
+| **Cohere** | Command R+, Command R, Aya | api.cohere.com |
+| **Perplexity** | Sonar Pro, Sonar, Sonar Reasoning | api.perplexity.ai |
+| **01.AI** | Yi-Large, Yi-Vision | api.01.ai |
+| **Alibaba** | Qwen-Max, Qwen-Plus, Qwen-Turbo | dashscope.aliyuncs.com |
+| **Baidu** | ERNIE 4.5, ERNIE Speed | aip.baidubce.com |
+| **Amazon Bedrock** | Claude, Titan, Llama, Mistral (via AWS) | bedrock.amazonaws.com |
+| **Azure OpenAI** | GPT-4.1, o3 (via Microsoft) | *.openai.azure.com |
+| **NVIDIA NIM** | Llama 3.3, Mistral, DeepSeek-R1 (on NVIDIA cloud) | integrate.api.nvidia.com |
+| **Together AI** | 50+ open models via API | api.together.xyz |
+| **Groq** | Llama, Mixtral, Gemma (ultra-fast inference) | api.groq.com |
+| **Fireworks AI** | Llama, Mixtral, DeepSeek | api.fireworks.ai |
+
+### Local AI
+
+Any local model is compatible — Halyn intercepts at the process level, not the network level.
+
+| Runtime | Models | Notes |
+|---------|--------|-------|
+| **Ollama** | Llama 3.3, Qwen2.5, Mistral, DeepSeek-R1, Phi-4, Gemma 3, ... | OpenAI-compatible API |
+| **LM Studio** | Any GGUF model | OpenAI-compatible server |
+| **Jan.ai** | Any GGUF or ONNX model | Desktop + server mode |
+| **GPT4All** | Llama, Mistral, Phi variants | Offline, no telemetry |
+| **llama.cpp** | Any GGUF model directly | Server mode (`--server`) |
+| **LocalAI** | 100+ models, any GGUF | Drop-in OpenAI replacement |
+| **text-generation-webui** | Any HuggingFace model | Extension ecosystem |
+| **KoboldCpp** | Any GGUF model | Focus on creative writing |
+| **OpenWebUI** | Ollama + OpenAI frontend | Browser-based |
+| **AnythingLLM** | Multi-model workspace | Team-friendly |
+| **Xinference** | HuggingFace + GGUF | Enterprise local inference |
+| **vLLM** | HuggingFace models | High-throughput server |
+| **TGI (HuggingFace)** | HuggingFace models | Production inference |
+
+### Agentic frameworks
+
+Halyn intercepts any agentic system. The agent framework doesn't matter.
+
+| Framework | Notes |
+|-----------|-------|
+| **OpenClaw** | Full interceptor — every action audited |
+| **Claude Cowork** | Proxy + filesystem hooks |
+| **Claude Code** | Process-level monitoring |
+| **LangChain** | API calls intercepted automatically |
+| **LlamaIndex** | API calls intercepted automatically |
+| **AutoGen** | API calls intercepted automatically |
+| **CrewAI** | API calls intercepted automatically |
+| **Semantic Kernel** | API calls intercepted automatically |
+| **BeeQ** | Native AAP integration |
+| **Any MCP agent** | MCP server passthrough |
+| **Any A2A agent** | Network-level interception |
+| **Any OpenAI-compatible API** | Universal proxy compatibility |
+
+### The rule
+
+> If an AI touches your machine or calls an API — Halyn sees it.
+> No exception. No exclusion. That's the point.
 
 ---
 

{halyn-2.1.1 → halyn-2.1.3}/tests/test_halyn.py

@@ -1,4 +1,7 @@
-# Copyright (c) 2026 Elmadani SALKA. All rights reserved.
+# Copyright (c) 2026 Elmadani SALKA
+# Licensed under BSL-1.1. See LICENSE file.
+# Commercial use requires a license — contact@halyn.dev
+
 """
 Halyn Test Suite — Every README example must pass.
 

halyn-2.1.1/src/halyn/llm.py

@@ -1,180 +0,0 @@
-# Copyright (c) 2026 Elmadani SALKA
-# Licensed under BSL-1.1. See LICENSE file.
-# Commercial use requires a license — contact@halyn.dev
-
-"""
-LLM Connector — Multi-provider LLM abstraction.
-
-Supports: Claude API, OpenAI API, Ollama (local), HuggingFace (local),
-          vLLM (self-hosted), any OpenAI-compatible endpoint.
-
-The LLM is NOT in the control plane. It connects FROM OUTSIDE via MCP.
-This module handles outbound LLM calls when Halyn needs reasoning
-(e.g. for autonomous reasoning, incident analysis, summarization).
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-import os
-from abc import ABC, abstractmethod
-from typing import Any
-
-log = logging.getLogger("halyn.llm")
-
-
-class LLMConnector(ABC):
-    """Base class for LLM connections."""
-
-    @abstractmethod
-    async def complete(self, prompt: str, system: str = "", max_tokens: int = 1000) -> str:
-        """Send prompt, get response."""
-
-    @abstractmethod
-    async def is_available(self) -> bool:
-        """Check if LLM is reachable."""
-
-
-class ClaudeConnector(LLMConnector):
-    """Anthropic Claude API."""
-
-    def __init__(self, api_key: str = "", model: str = "claude-sonnet-4-6") -> None:
-        self.api_key = api_key or os.environ.get("ANTHROPIC_API_KEY", "")
-        self.model = model
-        self.endpoint = "https://api.anthropic.com/v1/messages"
-
-    async def complete(self, prompt: str, system: str = "", max_tokens: int = 1000) -> str:
-        import aiohttp
-        headers = {
-            "x-api-key": self.api_key,
-            "anthropic-version": "2023-06-01",
-            "content-type": "application/json",
-        }
-        body: dict[str, Any] = {
-            "model": self.model,
-            "max_tokens": max_tokens,
-            "messages": [{"role": "user", "content": prompt}],
-        }
-        if system:
-            body["system"] = system
-        async with aiohttp.ClientSession() as session:
-            async with session.post(self.endpoint, json=body, headers=headers) as resp:
-                data = await resp.json()
-                return data.get("content", [{}])[0].get("text", "")
-
-    async def is_available(self) -> bool:
-        return bool(self.api_key)
-
-
-class OpenAIConnector(LLMConnector):
-    """OpenAI or any OpenAI-compatible API (vLLM, LiteLLM, etc.)."""
-
-    def __init__(self, api_key: str = "", model: str = "gpt-4.1",
-                 endpoint: str = "https://api.openai.com/v1") -> None:
-        self.api_key = api_key or os.environ.get("OPENAI_API_KEY", "")
-        self.model = model
-        self.endpoint = endpoint
-
-    async def complete(self, prompt: str, system: str = "", max_tokens: int = 1000) -> str:
-        import aiohttp
-        headers = {
-            "Authorization": f"Bearer {self.api_key}",
-            "Content-Type": "application/json",
-        }
-        messages = []
-        if system:
-            messages.append({"role": "system", "content": system})
-        messages.append({"role": "user", "content": prompt})
-        body = {"model": self.model, "max_tokens": max_tokens, "messages": messages}
-        async with aiohttp.ClientSession() as session:
-            async with session.post(f"{self.endpoint}/chat/completions",
-                                    json=body, headers=headers) as resp:
-                data = await resp.json()
-                return data.get("choices", [{}])[0].get("message", {}).get("content", "")
-
-    async def is_available(self) -> bool:
-        return bool(self.api_key)
-
-
-class OllamaConnector(LLMConnector):
-    """Ollama local inference. Zero cost, zero internet."""
-
-    def __init__(self, model: str = "llama3.2", host: str = "http://localhost:11434") -> None:
-        self.model = model
-        self.host = host
-
-    async def complete(self, prompt: str, system: str = "", max_tokens: int = 1000) -> str:
-        import aiohttp
-        body: dict[str, Any] = {
-            "model": self.model,
-            "prompt": prompt,
-            "stream": False,
-        }
-        if system:
-            body["system"] = system
-        async with aiohttp.ClientSession() as session:
-            async with session.post(f"{self.host}/api/generate", json=body) as resp:
-                data = await resp.json()
-                return data.get("response", "")
-
-    async def is_available(self) -> bool:
-        try:
-            import aiohttp
-            async with aiohttp.ClientSession() as session:
-                async with session.get(f"{self.host}/api/tags", timeout=aiohttp.ClientTimeout(total=3)) as resp:
-                    return resp.status == 200
-        except Exception:
-            return False
-
-
-class HuggingFaceConnector(LLMConnector):
-    """Run any HuggingFace model locally. Zero cloud."""
-
-    def __init__(self, model: str = "mistralai/Mistral-7B-Instruct-v0.3") -> None:
-        self.model_name = model
-        self._pipeline: Any = None
-
-    async def complete(self, prompt: str, system: str = "", max_tokens: int = 1000) -> str:
-        if self._pipeline is None:
-            self._load()
-        full_prompt = f"{system}
-
-{prompt}" if system else prompt
-        result = self._pipeline(full_prompt, max_new_tokens=max_tokens, do_sample=True, temperature=0.7)
-        return result[0]["generated_text"][len(full_prompt):]
-
-    async def is_available(self) -> bool:
-        try:
-            import transformers  # noqa: F401
-            return True
-        except ImportError:
-            return False
-
-    def _load(self) -> None:
-        from transformers import pipeline
-        log.info("llm.loading model=%s (this may take a while...)", self.model_name)
-        self._pipeline = pipeline("text-generation", model=self.model_name, device_map="auto")
-        log.info("llm.loaded model=%s", self.model_name)
-
-
-# ─── Factory ────────────────────────────────────────
-
-def create_connector(provider: str, **kwargs: Any) -> LLMConnector:
-    """Create an LLM connector by name."""
-    connectors: dict[str, type[LLMConnector]] = {
-        "claude": ClaudeConnector,
-        "anthropic": ClaudeConnector,
-        "openai": OpenAIConnector,
-        "gpt": OpenAIConnector,
-        "ollama": OllamaConnector,
-        "huggingface": HuggingFaceConnector,
-        "hf": HuggingFaceConnector,
-        "vllm": OpenAIConnector,  # vLLM is OpenAI-compatible
-        "litellm": OpenAIConnector,
-    }
-    cls = connectors.get(provider.lower())
-    if cls is None:
-        raise ValueError(f"Unknown LLM provider: {provider}. Available: {list(connectors.keys())}")
-    return cls(**kwargs)
-