Dexto: AI Agent Harness and Toolkit Released

What changed

Truffle AI has introduced Dexto, an open-source agent harness and toolkit designed for the development of AI agents and agentic applications. Dexto functions as an orchestration layer, enabling Large Language Models (LLMs) to act as reliable, stateful agents capable of performing actions, maintaining context, and recovering from errors. It is conceptualized as an operating system for AI agents, where the LLM serves as the CPU, the context window as RAM, Dexto as the operating system, and the developer's agent as the application.

The framework emphasizes a configuration-driven approach, allowing agents to be defined in YAML, which facilitates swapping models and tools without altering code. Key features include out-of-the-box session management, tool orchestration, memory capabilities, multimodal support, and observability. Dexto supports deployment across various environments, including local, cloud, and hybrid setups, and offers multiple interfaces such as a Command Line Interface (CLI), a Web User Interface (UI), a REST API, and integrations with platforms like Discord and Telegram.

A notable component is the production-ready coding agent that ships with Dexto. This agent can autonomously build, debug, and refactor code, execute shell commands, run tests, spawn specialized sub-agents, and maintain context across sessions using persistent memory. It is compatible with over 50 LLMs, allowing for model switching mid-conversation, and integrates with more than 30 tools via the Model Context Protocol (MCP), including Puppeteer, Linear, and ElevenLabs. Dexto also provides human-in-the-loop controls for fine-grained permissions and supports multi-agent systems where agents can spawn ephemeral sub-agents for complex tasks.

Installation and Usage

Installation can be done via a native installer for macOS/Linux/WSL or PowerShell for Windows, or by building from source. The CLI offers various modes, including interactive terminal interaction (dexto --mode cli), one-shot task execution (dexto -p "task"), and non-interactive modes for MCP integration. Users can configure defaults, such as LLM providers and API keys, using dexto setup.

Dexto Agents SDK

For programmatic development, Dexto provides an SDK (@dexto/core) that allows developers to build AI agents. The SDK supports creating agents, managing sessions, generating responses (including streaming and multimodal inputs), and switching LLMs. It also enables starting a Dexto server programmatically to expose REST and SSE streaming APIs for integration with external applications.

Why it matters for builders

Dexto offers a comprehensive framework that abstracts away much of the complexity involved in building sophisticated AI agents. Its declarative configuration, extensive LLM and tool support, and flexible deployment options empower developers to rapidly prototype and deploy agentic applications. The inclusion of a powerful coding agent and the ability to create multi-agent systems further lowers the barrier to entry for complex AI-driven workflows.

Practical impact

Developers can leverage Dexto to build custom coding assistants, autonomous task executors, or digital companions. The framework's support for MCP allows seamless integration with existing tools and services. Builders can start by installing the CLI and experimenting with the pre-built coding agent, or by using the SDK to programmatically define and manage their own agents. The ability to switch LLMs on the fly and manage persistent sessions provides a flexible development and user experience.

Caveats and source limits

The provided source material focuses on the capabilities and architecture of Dexto, including its CLI and SDK. Specific details regarding performance benchmarks, pricing for any potential enterprise offerings, or a definitive release date beyond the "fresh release" tag are not present. The source also mentions "6 AI signals, 5 developer signals" without elaborating on what these metrics represent. The latest release mentioned is @dexto/tui@1.16.6, but a comprehensive changelog or version history is not detailed.