Timbal

Timbal is an end-to-end AI platform built for enterprise teams, combining an open-source Python framework with a managed cloud platform (Timbal Studio) to build, deploy, and govern production AI agents, workflows, interfaces, and knowledge bases. The core framework is licensed under Apache 2.0 and available on GitHub, while the broader platform adds managed infrastructure, a visual Studio, enterprise security controls, and 100+ native integrations. Built in Barcelona, Timbal positions itself as a full-stack replacement for the fragmented toolchain of frameworks, workflow builders, vector stores, UI tools, and observability platforms that enterprise AI teams typically assemble.

What It Is

Timbal provides two complementary layers: an open-source Python/TypeScript framework for developers who want typed, async-native agent and workflow primitives, and a managed SaaS platform (Timbal Studio) for teams that need visual building, governed deployments, and enterprise integrations. Agents handle autonomous reasoning with tools and memory; Workflows define deterministic pipelines with explicit dependency resolution. Both share one calling interface and compose without glue code. The platform also ships a Hybrid DB engine (LanceDB + DuckDB) for vector, full-text, and SQL search in a single query, a TypeScript SDK, a CLI, and an MCP endpoint at api.timbal.ai/mcp.

The ACE Control Layer

A distinctive proprietary component is ACE (Action Control Engine), described by Timbal as a behavioral runtime that sits as a proxy in front of any LLM. According to the product page, ACE delivers a claimed +30% reliability gain and 0.1× cost per run versus baseline by steering smaller models with deterministic rules—schemas, refusals, tool-use policies, handoff logic—rather than prompt engineering. It is provider-agnostic and requires only a one-line base URL change to adopt with existing OpenAI or Anthropic SDK clients.

Developer-First Architecture

The framework is designed around a small, hackable core (under 10,000 lines per the README) with no hidden magic:

• Agents and Workflows share one event-stream interface; compose freely with MCP servers, custom tools, and ToolSets that resolve per-request based on session role or context.
• Memory compaction is automatic, with composable strategies (keep last N turns, summarize, compact tool results) triggered at a configurable context-window threshold.
• Human-in-the-loop gates pause any run, persist state, and resume across process restarts via parent_id.
• Evals are declarative YAML suites that validate output semantics, timing, token usage, tool inputs/outputs, and flow shape (sequences, parallel, nested patterns).
• HTTP serving turns any Agent or Workflow into a /run, /stream, /cancel server with auto-generated JSON schemas in one command.
• The framework publishes benchmark comparisons against LangGraph, CrewAI, OpenAI Agents SDK, PydanticAI, and Agno; the README reports p50 latency of 1.1 ms and throughput of 1,716 runs/second at concurrency 10 in overhead benchmarks with faked LLMs.

Enterprise Deployment and Governance

Timbal supports multi-tenant SaaS, dedicated VPC, and fully on-premises deployments across AWS, Azure, and GCP. Enterprise features include:

• SOC 2 Type II audit in progress, ISO 27001, GDPR, and EU AI Act alignment
• EU-region data hosting with GDPR-aligned DPAs
• SSO (SAML 2.0, OIDC) and SCIM provisioning for Okta, Azure AD, and Google Workspace
• Role-based access controls enforced at the UI, SDK, and API layers
• Audit logs, run-level traceability, and eval-gated promotion across dev/staging/prod environments
• Customer-owned model API keys; model-agnostic routing across OpenAI, Anthropic, Google, Mistral, Llama, and any OpenAI-compatible endpoint

Integrations and Stack Coverage

The platform ships 100+ native connectors including SAP, Salesforce, Slack, Microsoft Teams, Google Drive, Jira, HubSpot, Zendesk, Airtable, Stripe, GitHub, Asana, and major cloud providers. It also supports any MCP server over stdio or HTTP. The homepage positions Timbal as a replacement for infrastructure tools (Vercel, Railway, Docker), agentic frameworks (LangChain, CrewAI, Haystack), workflow builders (n8n, Zapier, Make), UI tools (Streamlit, Gradio, Next.js), vector stores (Pinecone, Weaviate, Chroma, Qdrant), LLM routers (LiteLLM, OpenRouter), and observability tools (LangSmith, Weights & Biases).

Update: v2.2.8

The latest GitHub release is v2.2.8, published on July 6, 2026. The repository was last pushed on July 6, 2026, and the Python framework is listed on PyPI. The developers page references the Python framework as v1.4.0 released April 2026, while the GitHub release tag reflects the more recent v2.2.8 cadence. The product roadmap lists Voice Agents and Timbal Workspace as upcoming products, and the homepage announces "Timbal Compose" as a new feature.