Stop harmful AI agent actions in real time. Two lines of code. MIT Licensed.

Why AeneasSoft?

• In-Process Circuit Breaker: Stops harmful agent actions in RAM before they execute. No proxy. No latency.

• Framework Agnostic: Works at the HTTP transport layer, below every AI framework. OpenAI, Anthropic, Gemini, Mistral, Groq, Cohere — automatic.

• EU AI Act Ready: Automated Article 12 compliance reports with RSA-2048 signatures.

Quickstart (30 seconds)

pip install aeneas-agentwatch

Every LLM call is now monitored. Rogue agents blocked in RAM.`

Read the full documentation →

Active Defense (Circuit Breaker)

The killer feature. Block runaway agents in application memory:

Per-agent budgets:

with agentwatch.agent("ExpensiveBot", budget_per_hour=5.0, block_on_threshold=True): result = client.chat.completions.create(...)

→ CircuitBreakerException if budget exceeded. Request never sent.`

Comparison

AeneasSoft Langfuse LangSmith Helicone

In-Process Blocking ✅ ❌ ❌ ❌

Circuit Breaker ✅ ❌ ❌ ❌

No Proxy Required ✅ ✅ ✅ ❌

Zero-Code Setup ✅ ❌ ❌ ❌

EU AI Act Reports ✅ ❌ ❌ ❌

Open Source (MIT) ✅ ✅ ❌ Partial

Architecture (Patent Pending)

Layer 1 wraps SDK methods for rich structured data. Layer 2 catches everything at HTTP level for framework-agnostic coverage. The deduplication flag (thread-local + ContextVar) ensures no duplicate spans.

Node.js SDK npm install @aeneassoft/sdk-node

Self-Hosting (Docker)

Verify it's running:

curl http://localhost:3001/health

Then connect the SDK:

SDK auto-connects to localhost:3001. No API key needed in local mode.`

Full Stack (production): docker compose up -d — adds Kafka + Proxy. Configure via .env.

Supported Providers

Works automatically with: OpenAI, Anthropic, Gemini, Mistral, Groq, Cohere, Together AI, Fireworks, Azure OpenAI, Ollama — and any provider accessible via HTTP.

Cost tracking for 20+ models with current list prices.

What We Don't Do (Yet)

We believe in transparency over hiding behind "Beta" labels.

• Prompt management / versioning — not our focus, use LangSmith or Humanloop for that

• A/B testing / evaluation pipelines — we observe and protect, we don't evaluate

• Streaming chunks: captures request + final usage summary, not individual chunks (full chunk-level tracing Q3 2026)

• Cost precision: list prices for 20+ models — batch API, cached tokens, and fine-tuned rates are not reflected

License

MIT License — the SDK, interceptor, circuit breaker, and dashboard are all open source.

The method (Dual-Layer Telemetry Interception and Active Defense) is protected by a USPTO Provisional Patent (April 2026).

If this is useful, please star this repo — it helps us reach more developers.

• Discord

• Website

• Docs

• PyPI

• npm