Orchestrate. Collaborate. Deliver.
A local-first desktop workbench for running and reviewing AI coding agents
against your own repositories.
Download · Source · Trust & Safety
Free and open source. macOS, Windows and Linux.
TaskWraith coordinates coding agents from multiple providers in one desktop app. Stay focused with one agent, isolate parallel work in sub-threads, or invite a small Ensemble to deliberate and review in a shared transcript.
- Every model, one thread. Bring compatible provider seats together without losing the conversation, workspace timeline, or review trail.
- Nothing lands without you. Give every seat explicit authority, then review TaskWraith-brokered approvals, activity, and diffs in one place.
- Your bench, your look. Shape the palette, density, layout, and Multiview workspace around the way you work.
- Spend with no surprises. Keep provider usage signals and run activity visible while each account and quota stays under its provider's control.
- Powerful on day one. Welcome, Goals, Diff Studio, Canvas, terminal, and familiar chat workflows are ready when you need them.
 Set up your bench |
 Keep the work in view |
 Review before you land |
- Download TaskWraith from the
latest GitHub release.
Choose the checksum-listed installer for macOS, Windows, or Linux and verify
it against the release's
SHA256SUMS-<version>.txt. - Bring a provider. Install and sign in to the provider CLI or account you want to use; TaskWraith does not bundle those credentials or bypass provider authentication. Local Ollama models are supported too.
- Begin somewhere safe. Add a scratch repository, start with Ask or a read-only posture, and widen access only after you are comfortable reviewing approvals, activity, and diffs.
The optional iPhone and iPad companion is in TestFlight beta and pairs with TaskWraith on macOS; it is not a standalone AI app.
For contributors, install Node.js 20 or newer and any provider CLI you plan to use, then run:
npm ci
npm run devSource-ahead note: this checkout may contain work newer than the latest public tag. The top
CHANGELOG.mdentry describes source-ahead changes; they are not shipped until a new tag and matching artifacts are published.
| Explore | What you will find |
|---|---|
| Positioning | The product promise, claim boundaries, and a guide to solo chats, sub-threads, and Ensembles |
| Trust & Safety | Safe first runs, permissions, storage, provider boundaries, and release verification |
| Privacy | What stays local, what providers receive, optional observation, and remote companion boundaries |
| Security | Supported versions, reporting, disclosure, and release security |
| Safety | Engineering guardrails, implemented protections, and known limits |
| Model Catalogue | Current provider and model rows, reasoning controls, and route-specific limits |
| Changelog | Shipped releases and the clearly labelled source-ahead section |
| Advanced setup | Optional accounts, services, remote features, and platform permissions |
| Architecture | The desktop, runtime, orchestration, and persistence map |
TaskWraith keeps your workspace, local history, and approvals on your machine. Prompts and run context go to whichever provider you select. Provider tools and native boundaries vary, so use the catalogue and trust guide for the current details.
TaskWraith is Apache-2.0 licensed. Provider product names identify compatible integrations and do not imply endorsement; provenance and source-specific rights notes live with the relevant design assets and in Terms Notes.