Dexto review, pricing and limits

Dexto replaces building your own agent infrastructure from scratch: developers get a working coding agent, memory, and tool calling out of the box instead of wiring it together by hand. Its npm installs sit around 1,100 for the past month, a modest but real number next to venture-backed coding-agent rivals.

Dexto is an open-source agent framework from Truffle AI that had drawn 648 GitHub stars by August 2026, giving developers a configurable orchestration layer for building AI agents instead of a single chatbot. It ships a ready-to-run coding agent that plans multi-step tasks and calls tools through the Model Context Protocol, switching freely between language models from any provider.

Maker: Truffle AI · Autonomy: semi autonomous · Maturity: BETA

Underlying models: OpenAI gpt-5.2 family (gpt-5.2, gpt-5.2-pro, gpt-5.2-codex, o4-mini), Anthropic Claude Sonnet, Opus and Haiku, Google Gemini 3 Pro and Gemini 2.5 Pro/Flash, Groq-hosted Llama 4, Qwen and DeepSeek, xAI Grok 4 and Grok 3, Local models via Ollama/GGUF (Llama, Qwen, Mistral)

About Dexto

Dexto is the open agent orchestration framework built by Truffle AI, a two-person team in Bengaluru, India that raised a $500,000 seed round from Y Combinator as part of the February 2025 Winter 2025 batch. The project began as an internal framework called Saiki before Truffle AI rebuilt and renamed it Dexto, open-sourcing the layer that previously only powered the company's own hosted agent APIs. Where a typical AI chatbot answers one prompt at a time, Dexto is built to hold state across a whole task: it plans, calls tools, checks its own output, and picks up where it left off after a restart. Dexto runs as a local CLI or a Web UI and connects to a wide range of large language models across seven providers, from closed frontier APIs to fully local models run through Ollama, letting a user swap between them mid-conversation without restarting a session. It ships with a coding agent by default that reads a codebase, writes and refactors files, runs shell commands and tests, and can delegate exploration work to a dedicated sub-agent before making a change. Every agent reaches outside tools through the Model Context Protocol, with a built-in store of prebuilt connectors covering browser control, issue tracking, and voice generation. Because agents are defined in a single YAML file rather than custom code, Dexto is aimed at developers who want to build a purpose-built agent, such as a support bot on WhatsApp or an analytics assistant in Slack, the two examples Truffle AI cites from its own hosted-API product, without writing the state management, memory, and vector-database plumbing themselves. Its permission system defaults to asking approval before each tool call, with an auto-approve mode for developers who trust it to work unattended on their own machine. Dexto's source is published on GitHub and the CLI installs directly from dexto.ai, so running it costs nothing beyond whichever LLM provider or local GPU a user chooses to power it with. The project ships new CLI and plugin package versions on a near-weekly cadence: dexto reached version 1.10.4 on August 18, 2026, and its GitHub repository, created in March 2025, had logged 648 stars and 79 forks by that same release. npm records roughly 1,100 downloads of the dexto package over the month ending August 29, 2026, a small but real signal for a two-person team's open-source project.

Pricing

Dexto itself has no price: it is free to self-host under the source-available Elastic License 2.0, and the only running cost is whichever LLM provider or local model a user connects. Truffle AI's separate hosted-API product at trytruffle.ai runs its own freemium plan for teams that prefer a managed API instead of self-hosting.

Key Features

• 50+ Swappable LLMs: Switches between OpenAI's gpt-5.2 family, Anthropic's Claude Sonnet, Opus and Haiku, Google's Gemini 3 Pro, xAI's Grok 4, and local Ollama models mid-conversation with no restart.
• Built-in Coding Agent: Ships a default agent that reads, writes, and refactors a full codebase, executes shell commands, and runs tests from the CLI or a Web UI.
• MCP Store With 30+ Tools: Connects to Model Context Protocol servers including Puppeteer, Linear, ElevenLabs, and Firecrawl through a built-in browsable store.
• Sub-Agent Delegation: The coding agent hands off exploration work to a dedicated read-only helper agent before it edits any code, then auto-cleans that helper up once it finishes.
• Human-in-the-Loop Permissions: Defaults to asking approval before every tool call, remembers approvals per session, and offers an auto-approve mode for trusted local runs.
• Persistent Sessions: Conversations and memories persist across restarts, and a built-in search command finds prior sessions by keyword.

Strengths

• No subscription cost: Dexto runs free of charge on your own machine, unlike managed coding-agent subscriptions that start at $17 to $20 a month for comparable functionality.
• Backed by an actively maintained open-source project: 79 forks and a near-weekly release cadence point to real external usage beyond Truffle AI's own founding team.
• Ships sub-agent delegation and a Model Context Protocol tool store out of the box, functionality that closed rivals like Claude Code require separate MCP server setup to match.
• Actively developed: version 1.10.4 shipped August 18, 2026, with 68 open GitHub issues signaling day-to-day usage rather than an abandoned repository.

Weaknesses

• Small install base so far, per npm's own public download counter, well behind established coding-agent tooling.
• A two-person team runs the project, per public reporting, which limits how fast issues get triaged compared with venture-scale rivals that staff dedicated support.
• Source-available rather than a fully permissive open-source license like OpenClaw's Apache License, so reselling Dexto itself as a hosted service is restricted.

Visit Dexto Official Site