# TaskWing > TaskWing is an open source harness for coding agents (AGPL-3.0). It runs the coding agents you already use, each in its own sandbox, in parallel, and gives you back one draft pull request that has passed a hidden exam. It is free and self-hosted. How a run works: your own coding agent (the lead) plans the task and splits it into pieces with no shared files. A separate agent session writes a hidden exam of plain-sentence checks and tests, and you approve it before any agent starts. Each piece is built by an agent in its own sandbox. The work is joined on one branch, the exam runs in a clean sandbox, and the run ends as one draft pull request that you review and merge. Status (November 2026): proof of concept. The Go CLI runs `taskwing run`, `taskwing hunt` and `taskwing doctor`. Agents: Claude Code today; Codex CLI or OpenCode is next. Sandboxes: boat today, plus local folders for trying the flow; E2B, AWS and Docker are planned. A desktop app for macOS, Linux and Windows draws runs; it uses a demo engine until the engine stream ships. ## Docs - [Home page](https://taskwing.app/): what TaskWing does, the hidden exam, real runs and an FAQ - [About](https://taskwing.app/about/): who builds TaskWing, the licence, status and how to get in touch - [README](https://github.com/josephgoksu/TaskWing/blob/main/README.md): install and first run - [Proof of concept guide](https://github.com/josephgoksu/TaskWing/blob/main/docs/poc.md): setup, flags and the bug hunt - [Architecture](https://github.com/josephgoksu/TaskWing/blob/main/docs/ARCHITECTURE.md): every concept and how they connect - [Features](https://github.com/josephgoksu/TaskWing/blob/main/docs/FEATURES.md): what is built, partial and planned - [Principles](https://github.com/josephgoksu/TaskWing/blob/main/PRINCIPLES.md): the rules every feature is built on ## Use cases - [Build one feature with parallel coding agents](https://taskwing.app/use-cases/parallel-feature-work/): Give TaskWing one feature. Your coding agent splits it into pieces with no shared files, agents build them in parallel sandboxes, a hidden exam checks it. - [Find bugs and prove each one with a failing test](https://taskwing.app/use-cases/bug-hunt/): taskwing hunt writes a test for every bug it suspects, keeps only tests that fail on today's code, then fixes each proven bug in its own sandbox. One pull request. - [Run coding agents in sandboxes with no GitHub access](https://taskwing.app/use-cases/sandboxed-agents/): TaskWing runs each coding agent in its own sandbox. Code goes in as a git bundle, only commits come out, and your machine pushes. What is enforced today. ## Compare - [TaskWing vs Conductor](https://taskwing.app/compare/conductor/): Conductor is a Mac app for steering Claude Code and Codex sessions in git worktrees. TaskWing is an open source harness: one task in, one examined PR out. - [TaskWing vs Vibe Kanban](https://taskwing.app/compare/vibe-kanban/): Vibe Kanban is an open source board that runs 10+ coding agents in local git worktrees. TaskWing turns one task into one PR checked by a hidden exam. - [TaskWing vs Claude Squad](https://taskwing.app/compare/claude-squad/): Claude Squad runs Claude Code, Codex, Gemini and Aider in tmux sessions and git worktrees. How it differs from TaskWing's sandboxed, examined runs. - [TaskWing vs GitHub Copilot cloud agent](https://taskwing.app/compare/github-copilot-cloud-agent/): GitHub Copilot cloud agent (formerly coding agent) opens a PR from GitHub's cloud. TaskWing is an open source harness that runs your own agents in parallel. ## Blog - [What is an agent harness?](https://taskwing.app/blog/what-is-an-agent-harness/): An agent harness is the software around coding agents: it splits work, isolates each agent, limits access, checks the result and hands it back. - [Why the agents never see the tests](https://taskwing.app/blog/hidden-exam/): An agent that writes its own tests can pass its own exam. TaskWing has a separate session write hidden tests first. How it works, and the run where it failed. - [A bug hunt on Osprey: 4 bugs proven, 2 shipped](https://taskwing.app/blog/osprey-bug-hunt/): taskwing hunt on Osprey, an open source Go engine: 4 bugs proven by failing tests and fixed in 8 min 58 s. 2 were already fixed on main. The receipt. ## Source - [GitHub repository](https://github.com/josephgoksu/TaskWing): AGPL-3.0 - [Desktop app](https://github.com/josephgoksu/TaskWing/tree/main/app): Tauri, plain HTML, CSS and JavaScript ## Optional - [Full text of every use case, comparison and blog post](https://taskwing.app/llms-full.txt)