Launching Draht: An AI Coding Agent Framework

by Oskar Freye

Draht is live — a modular AI coding agent with extensions, skills, multi-model support, and a custom TUI. Here is why I built it.

Draht AI Coding Agent

Two days ago, on February 28th, I publicly launched draht — Dynamic Routing for Agent & Task Handling. It’s a modular AI coding agent framework, and it’s the tool I’ve been wanting to exist for the past year.

Why Build Your Own?

I’ve been using AI coding tools daily — Copilot, Cursor, Claude. They’re great, but they all share the same problem: they’re closed systems. Want to add a custom tool? Switch models mid-session? Run it headless in CI? Integrate it with your Slack? Good luck.

Draht fixes this with a layered architecture where everything is pluggable:

The Extension System

This is where it gets interesting. Extensions can register tools, commands, providers, themes, skills, and prompt templates. The agent discovers them from your project config, and everything composes cleanly.

We just shipped GSD (Get Stuff Done) integration — a structured workflow with domain modeling, codebase mapping, and task planning, all driven through slash commands like /create-plan and /commit-task. It uses TDD by default: write the test, then the code.

Subagents

The latest feature: subagent delegation. The main agent can spawn specialized subagents for parallel execution. An architect agent designs, an implementer builds, a reviewer checks — all orchestrated from a single session.

What’s Next

The velocity is wild — we’re shipping daily releases (v2026.3.5, v2026.3.6 just this week). SST v4 migration is done, model pricing is updated, and the community is growing. @draht/mom (Slack bot), @draht/pods (GPU pod management), and @draht/web-ui are all in active development.

Install it: bun add -g @draht/coding-agent or curl -fsSL https://draht.dev/install.sh | bash

The future of coding is agents that work with your workflow, not against it. Draht is my bet on that future.