Dexto

Why it's ranked here

It earns a look for how deliberately layered the codebase is: business logic, HTTP routes, approval handling and the command-line entry point each live in separate packages with typed errors and schema-validated configuration rather than one sprawling script. The interface support is broad without being a thin wrapper: a documented REST API generated from route definitions, a command-line tool, and the ability to run as its own protocol server for other agent tools to call into. The license is Elastic 2.0, not a permissive open-source license, so anyone planning to resell this as a hosted service should read the terms before building on it.

What's good

The approval system is genuinely fine-grained: policies can require confirmation for every tool call, auto-approve for trusted local work, or allow-list specific actions by name, and the tool remembers what a person has already approved within a session so it does not re-ask. Persistent memory and session history mean a conversation can be resumed, searched, or handed to a different model mid-stream without losing context. The sub-agent system lets a coding assistant delegate a bounded, read-only exploration task to a helper that cleans itself up afterward, which keeps a long task from ballooning into one giant untrimmed context window.

Tradeoffs

The license is Elastic 2.0, not a standard open-source license: it blocks anyone from hosting Dexto itself as a competing managed service, which matters for a team hoping to build directly on top of it. Running models locally depends on the machine's own GPU being detected correctly, trading a cloud provider's reliability for setup work that varies by hardware. The codebase spans a dozen or more packages for the command line, web interface, server, registry and individual tool integrations: real separation of concerns, but more ground for a newcomer to cover before a first meaningful change.

How to use it well

This fits a developer who wants one coding assistant they can point at different models by cost or capability, instead of being locked to a single vendor. Start in manual approval mode to see exactly what it wants to run before trusting auto-approve on a real codebase. Wire it in through its own protocol-server mode rather than the chat interface when the goal is plugging it into another agent's toolchain, not using it directly. It is not the right pick for someone who wants a finished product with one fixed model and nothing to configure, since real use starts by writing a configuration file by hand.

Observed

License

Elastic License 2.0 (ELv2), a source-available license (not OSI-approved open source) that bars offering the software to third parties as a hosted or managed service.

Language

TypeScript, structured as a pnpm-only monorepo (an only-allow pnpm preinstall guard) built with Turborepo.

Interfaces

a CLI binary, a web UI, a REST/SSE API (Hono, with OpenAPI-generated documentation), and an MCP server mode over stdio or HTTP/SSE transport.

Packaging

Also distributed as an installable SDK package exposing an agent class with session, streaming, and multimodal generation methods for embedding in other Node.js applications.

Structural

Configuration is YAML-based: models, tools, and agent behavior are declared in config files rather than in code.

Structural

Validation is schema-first via Zod across the core package, with a documented convention of explicit input/output type extraction rather than a common shortcut inference pattern.

Structural

Test suite uses Vitest with separate unit and integration test configurations.

Structural

Core package defines two typed error base classes, one for single failures and one for multi-issue validation, instead of throwing plain Error objects.