torque

Torque

Torque is a local agent-orchestration workspace built on top of Claude Code and Codex. Chat with an orchestrator agent, run coding agents in isolated git worktrees, dispatch work from a kanban board, and watch tasks ship — from a native desktop window or your browser. Because Torque is a harness over the CLIs you already use, all the work runs through your existing Claude and Codex subscription plans.

Why Torque

Coding agents are powerful but messy in practice. Each one wants its own session, its own branch, its own context window, and its own terminal history. Spinning up three at once can quickly turn into three terminals, three worktrees, three half-remembered prompts, and a lot of "where was I?" friction.

Torque is a harness on top of Claude Code and Codex: it drives the CLIs you already run, so every agent uses your existing subscription plans — no extra API keys or per-token billing. I built it because the AI companies would rather lock me into their harnesses and their idea of how I should work. Torque is how I found to get the full value out of the subscriptions I already pay for, on my own terms, without conforming to a vendor's prescribed workflow.

How it works

Torque is a hierarchy of agents, each minding its own role. Workers write the code. Engineers sit above them: they orchestrate tasks across their workers and merge the results, serializing changes so conflicts never happen in the first place. That orchestration keeps engineers busy and their context precious — too busy to also plan at a high level. So that job falls to the architect, which does the high-level planning and is usually the role you want to talk to. It interfaces with the engineers and shields their context from human interruption; because it is typically less loaded than the engineers, it is the natural entrypoint to the whole system — though you can still drop down and talk to an engineer directly when you need to.

You drive everything from the chat with the architect. This matters because the underlying Claude Code or Codex session is busy almost constantly — fielding messages from other agents and digests from the system — so a message typed straight at it would just get buried in its context. Instead, the chat works both ways: your messages go in through the harness — the normal Claude Code or Codex input channel — while the architect replies back through an MCP tool call, and it keeps working in between. You read its replies and decide what to do next on your own time, without ever stalling the agent or polluting its working context. Meanwhile the kanban board shows what every agent is doing at all times — every task, who owns it, and where it sits in the pipeline.

What you get:

• A chat-first workspace where you talk to an architect that plans the work and runs the engineers and workers for you — backed by Claude Code and Codex on your own subscription plans.
• Groups that act like separate projects — each one fully isolated from the others — so you can run Torque across all your projects from a single native desktop window or browser tab.
• Automatic git worktrees per agent, with checkpointing, diff tracking, and engineer-driven merges back to your base branch.
• A built-in kanban board with lanes, drag-and-drop, derived subtasks, and human-review gates.
• One-click task dispatch: pick a task, pick an action, and Torque spawns an agent in a fresh worktree with the rendered prompt already sent.
• An optional embedded engineer agent that orchestrates dispatch, review, and merge across an entire group without you babysitting each worker.
• Reusable action templates with Jinja-rendered prompts and pipeline transitions.
• A torque CLI for scripting from the command line.

Core Ideas

Torque builds Torque

Torque is built with Torque. Nearly all of its code is written by agents dispatched from the board and merged through the engineer — the same loop the product exposes to you. The agents produce code so fast that, as a deliberate experiment in trusting the orchestration loop, none of the code Torque has written for itself has been manually inspected line-by-line. Review, testing, and merge gates are themselves run by agents. It is an honest, sometimes uncomfortable, demonstration of what the harness can do — and a live stress test of the guardrails it ships with.

Agents only see what they're allowed to