context-labs/whip

A coding agent built for work that doesn’t fit in one context window.

Quickstart · What is this? · How it works · Benchmarks · Development

Download Desktop Beta — Apple Silicon · macOS 14 or newer.

Use a release built after the clean-project reset; old Desktop releases are not an upgrade path.

1. Download whipcode-desktop-darwin-arm64.dmg, open it, and drag Whip Beta into Applications.
2. Launch it and choose Set up this Mac if prompted. The app includes its matching backend; no separate CLI installation is needed.
3. Connect a model provider, open your project, and describe what you want to do.

Bring an API key or use a supported subscription login. Provider access is separate from installing WhipCode. See providers and authentication.

Prefer the terminal? Install the standalone CLI on macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/context-labs/whip/main/install.sh | sh
whipcode

For an exact version, use the release's install.sh command from its notes—no version environment variable is needed for releases with the two-script split. The companion latest.sh always selects stable v1+, never an alpha.

The new CLI track starts at v1.0.0-alpha.N. Prerelease installation is explicit; default stable installation becomes available with v1.0.0. Existing internal installs must follow the manual reset checklist, not upgrade in place.

Desktop-managed installations update through Desktop; standalone CLI installations use whipcode update. Setup and upgrades →

For the web client, start the daemon with whipcode daemon start, then run whipcode web. It opens your browser and stays running as a separate gateway; Ctrl+C stops web access, not daemon work. Ordinary daemon and Desktop startup open no web listener. See web access for --no-open, managed startup, and trusted-network configuration.

WhipCode is an open-source coding agent with a Desktop app, terminal UI, and web client. A shared daemon owns the work, so closing a window doesn’t end the session.

• Delegate recursively. Agents split work into focused tasks, launch their own agents, and coordinate through messages.
• Keep context within reach. Search and read history, files, and large tool outputs in bounded slices instead of cramming everything into a prompt.
• See what’s happening. Follow conversations, inspect the execution tree and timeline, and track model usage and cost.
• Stay in control. Choose models, set tool permissions, and connect local or remote execution hosts. Extend agents with skills and MCP servers.

1. Describe the task. Give an agent a goal and the context it needs.
2. Let it work. A recursive language model (RLM) loop uses short Starlark or JavaScript programs to inspect context, call tools, and delegate—not just a growing chat transcript.
3. Inspect and steer. Follow the trace, send another instruction, or reconnect later. The daemon keeps session state independent of the UI.

Explore the runtime →

WhipCode solved 20 of 30 tasks (66.7%) in a retained Frontier evaluation: 21 Terminal-Bench tasks and nine code-repository tasks.

Read the comparison with context: providers and execution environments differ. The supplied chart’s $2.60 WhipCode point and cost-axis definition are not yet reconciled with the retained reports; this is not a matched cost comparison or an official leaderboard ranking. Results, sources, and limitations →

Requires Go 1.27+, Node.js 24, and Task.

git clone --branch main https://github.com/context-labs/whip.git
cd whip
npm ci
task build

This builds locally without replacing an installed app or restarting its daemon. See the development and setup guide, Desktop guide, and contributor checks.

Documentation · Architecture · Frontend · Evaluations

Apache-2.0. Issues and pull requests are welcome—start with CONTRIBUTING.md.