Alquimia

Turn probabilistic AI into deterministic, governable enterprise infrastructure.

Get Started Runtime API Reference Core SDK Reference

Why Alquimia?

Most AI projects fail in production — not because the models are wrong, but because the infrastructure around them isn’t built for enterprise requirements: auditability, governance, security, and integration with existing systems.

Alquimia is the AI execution layer that closes this gap. It wraps LLMs and autonomous agents in explicit contracts, deterministic execution paths, and enterprise-grade integrations — so your AI investments deliver measurable outcomes, not just demos.

Ships as two open-source components:

alquimia-core — Python SDK. The agent execution engine: evaluation strategies, memory management, tool integration (MCP, Llama Stack, A2A), and OCI-based agent registry.
alquimia-runtime — FastAPI service. HTTP API wrapping the core SDK with async inference, SSE streaming, Redis-backed session state, Keycloak auth, Vault secrets, and a full registry management API.

Auditable by design

Every agent action is a typed, logged CloudEvent. Full execution traces for compliance, debugging, and replay — no black boxes.

Enterprise security

Keycloak OIDC, JWT, or API token auth. HashiCorp Vault for secrets. OCI artifact distribution for agent configs. Zero plaintext credentials.

Multi-model, multi-tool

Connect OpenAI, Anthropic, Ollama, vLLM, or any OpenAI-compatible endpoint. Integrate MCP servers, Llama Stack, or other agents as tools.

Production-ready infrastructure

Kubernetes-native with Kustomize overlays. Kafka-backed event bus with HMAC-signed CloudEvents. OpenTelemetry observability. PostgreSQL audit logs. Redis session state.

Built for enterprise evaluation

Governance & control

Agent configurations are versioned, validated, and distributed as OCI artifacts. Hold/unhold agents without redeployment. Dry-run validation before publishing.

Human-in-the-loop

Require human approval before any tool execution. Configurable per tool: NONE, LAZY, or REQUIRED. Approval flows over the same SSE stream.

Multi-channel

Agents receive messages from WhatsApp, Slack, Email, and web — normalized through the same execution pipeline. One agent, many channels.

Memory & context

Short-term token windowing and long-term Chain-of-Density summarization. Session state persisted in Redis with configurable TTL. RAG via Qdrant.

Explore the docs

Introduction What Alquimia solves, how it compares, and when to use it.

Architecture Event-driven execution model, CloudEvents protocol, and service topology.

Runtime API Reference Inference, streaming, context, registry, and knowledge base endpoints.

Core SDK Reference evaluate(), AssistantConfig, evaluation strategies, memory, tools.

Source repositories

| Repository | Description | |---|---| | Alquimia-ai/alquimia-core | Python SDK — agent execution engine, evaluation strategies, memory, tools, registry | | Alquimia-ai/alquimia-runtime | FastAPI service — HTTP API, async inference, SSE streaming, Keycloak auth, Vault secrets |