// the find
truffle-ai/dexto
A coding agent and general agent harness for building and orchestrating agentic applications.
Dexto is a TypeScript agent harness that wraps LLMs with session management, MCP tool integration, and multi-agent coordination — positioned as the 'OS for AI agents'. Ships with a coding agent out of the box and lets you define custom agents in YAML. Aimed at developers who want a Claude Code-style coding agent they can embed, extend, or self-host.
YAML-based agent definitions are genuinely portable — version-controlled configs that swap models and tools without code changes is a practical win over hardcoding. MCP support is first-class and bidirectional: Dexto can consume MCP servers or expose itself as one, which means it slots into Claude Code or Cursor as a tool rather than competing with them. The multi-backend storage story (Redis, Postgres, SQLite, in-memory) means you can actually run this in production without immediately hitting persistence walls. Sub-agent spawning with automatic cleanup and unified approval flow is a real architectural detail, not just a bullet point — the explore-agent pattern (read-only fast sub-agent) shows they've thought about the cost/speed tradeoff.
Elastic License 2.0 is not open source — you cannot use this to build a competing SaaS product, which the 'open source' framing in the README obscures. The model list in the README references models like 'gpt-5.2', 'gemini-3-pro', and 'grok-4' that don't exist as of mid-2026, which suggests the docs are aspirational or AI-generated rather than tested. At 631 stars and beta status, the production story is unproven — the coding agent is essentially reimplementing what Claude Code already does, and the differentiation over just using Claude Code's MCP server mode directly is thin. Anonymous telemetry is on by default with no install-time opt-out prompt, which will bite anyone deploying this in a corporate environment.