Donkit is an orchestration and deployment platform that acts as a bridge between high-level agent descriptions and low-level execution. They occupy the "RAGOps" and infrastructure layer of the agent stack, providing the necessary plumbing—memory, typed logic, and UI generation—that builders would otherwise have to construct manually. Their focus on "typed" agents is a significant push toward determinism in an ecosystem often criticized for being unpredictable.

For the broader ecosystem, Donkit matters because it addresses the deployment bottleneck. By automating the creation of custom interfaces and providing one-button hosting (both cloud and on-prem), they lower the barrier for non-technical or semi-technical users to ship functional agents. They are championing a shift from agents as experiment-driven prompts to agents as structured, observable software components.

The typed graph architecture

Software development has historically required a translation layer between the person who understands a business problem and the engineer who writes the code. Donkit operates on the premise that LLMs can now act as this translation layer, provided the platform underneath is rigorous. They describe this process as "vibecoding," but the technical reality is structured around a typed graph engine. In most agent frameworks, prompts are passed as loose strings, a method that frequently breaks in production as model outputs vary. Donkit instead types every node in the agent's logic flow. Every tool call and data transfer is validated against a schema before it runs, bringing the predictability of traditional software engineering to agentic workflows.

From prompt to production UI

One of the primary friction points in agent adoption is the interface. Most agents are confined to a simple chat box, but real-world tasks often require structured data visualization or specific input forms. Donkit uses a meta-agent to interpret a user's natural language intent—such as "build a support agent that reads Notion and replies on Slack"—and then generates not just the logic, but a custom user interface. This includes dashboards, tables, and charts tailored to the agent's specific task. This approach moves agents away from being simple chatbots and toward being functional micro-SaaS applications that can be shared with a team in one click.

Memory and multi-model flexibility

Reliability in agents depends heavily on how they handle context over time. Donkit implements a dual memory system: vector memory for broad knowledge retrieval and episodic memory for session-specific context. This is managed on a per-user and per-organization basis, ensuring data isolation. Architecturally, the platform is model-agnostic. Users can swap between models like Claude 3.5 Sonnet, GPT-4o, or Gemini mid-flight without rewriting the underlying agent logic. This flexibility is paired with token-level observability, allowing developers to trace the exact prompt, tool call, and latency of every step in a sequence to debug and optimize performance.

Origins in Tel Aviv and San Francisco

Founded in 2024, Donkit is led by a team of Alchemist Accelerator alumni with a background in scaling enterprise-grade B2B and B2C products. The company maintains a split operation: Tel Aviv serves as the hub for research and R&D, while San Francisco handles go-to-market strategy and partnerships. The team includes researchers and infrastructure engineers who focus on the "Day 2" problems of AI—the issues that arise only after an agent is deployed to real users. They are backed by major industry programs including NVIDIA Inception, AWS, and Microsoft for Startups, signaling an intent to serve the infrastructure needs of the growing agentic ecosystem. Their focus remains on converting "weekend demos" into software that can pass a corporate security review and handle revenue-moving tasks for household-name retailers.