Future AGI

Future AGI is a platform for evaluating, tracing, simulating, guarding, routing, and optimising AI agents in one feedback loop: simulate edge cases before launch, evaluate production behaviour, protect users in real time, and feed traces back into the next version. It is released under an open source core that can be self-hosted with Docker Compose, or used as a managed cloud service, and every layer is reachable through open interfaces: OpenTelemetry OTLP for traces, OpenAI-compatible HTTP for the gateway, and Postgres/ClickHouse SQL for storage. The current release is a nightly build for early testing.

Tracing is added by calling register(project_name=...) and instrumenting a client, for example OpenAIInstrumentor().instrument() in Python or new OpenAIInstrumentation().instrument() in TypeScript, after which existing calls are traced. The platform runs on Python 3.11+ (Django 5.1 with Channels), a Go gateway, and a React frontend, storing metadata in PostgreSQL, spans and time series in ClickHouse, state in Redis, and jobs in RabbitMQ with Temporal.

Client libraries are packaged separately: traceAI (OpenTelemetry instrumentation for Python, TypeScript, Java, and C#), ai-evaluation (metrics and guardrail scanners for Python and TypeScript), the futureagi platform SDK (datasets, prompts, knowledge base, experiments), agent-opt (prompt optimisation), simulate-sdk (voice-agent simulation via LiveKit and Silero VAD), and agentcc (gateway clients for Python, TypeScript, LangChain, LlamaIndex, React, and Vercel).

Features

• Simulate: thousands of multi-turn conversations against personas, adversarial inputs, and edge cases, for text and voice agents (LiveKit, VAPI, Retell, Pipecat)
• Evaluate: 50+ metrics under one evaluate() call, including groundedness, hallucination, tool-use correctness, PII, tone, and custom rubrics, combining LLM-as-judge, heuristic, and ML methods
• Protect: 18 built-in scanners (PII, jailbreak, injection, and others) and 15 vendor adapters such as Lakera, Presidio, and Llama Guard, inline in the gateway or as a standalone SDK
• Monitor: OpenTelemetry-native tracing across 50+ frameworks with span graphs, latency, token cost, and live dashboards
• Agent Command Center: an OpenAI-compatible gateway with 100+ providers, 15 routing strategies, semantic caching, virtual keys, MCP, and A2A; published benchmarks report about 29k req/s and P99 latency of 21 ms or less with guardrails on
• Optimise: six prompt-optimisation algorithms (GEPA, PromptWizard, ProTeGi, Bayesian, Meta-Prompt, Random) fed by production traces
• Integrations: LLM providers including OpenAI, Anthropic, Gemini, Bedrock, Azure, Mistral, Groq, and self-hosted Ollama, vLLM, and LM Studio; frameworks including LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, PydanticAI, DSPy, and Haystack; vector databases including Pinecone, Weaviate, Chroma, Milvus, Qdrant, and pgvector
• Deployment: Docker Compose with published images, a production overlay that generates secrets and pins image tags, cloud VMs, and air-gapped installs; Kubernetes/Helm manifests are planned