Build and automate with your choice of models.
Codewhale is an open-source agent that can read a project, edit files, run commands, and use connected tools. Direct the work from your terminal or local browser, choose the model, and keep the conversation and tool results in a saved session.
Models, teamwork, and control
Choose your models
Connect a provider with your own API key or a supported sign-in flow, use a gateway, or run a local model server. Choose the model for each session.
Share the work
Give parts of a larger task to agents with different models and roles. Save the team as a Fleet to use again.
Set the permissions
Choose the session mode and approval settings. Inspect tool calls and results, review file changes, interrupt work, and continue from a saved session.
Use the tools that fit your work
Tools let the agent act on a task: edit files, run a command, or use a connected service. Codewhale runs the agent and its tools locally, keeping the conversation and tool results together. Plugins use the permissions you grant.
- Files and commandsRead a project, edit files, run tests, and inspect command output within the permissions you set.
- Plugins and MCPConnect additional tools and services. Review and enable plugins before the agent can use them.
- Computer Use · previewA plugin that lets the agent see and operate other applications. Enable it explicitly and grant the system permissions it asks for.
- Saved sessionsKeep the conversation and tool results together. The local browser connects to the same Codewhale session on your computer; resume saved work instead of starting over.
- FleetAssign parts of a task to agents with different models and roles, and follow their progress.
Choose how your agents work
Use Plan to explore, Work to make changes, and Operate to coordinate agents. Approval settings control when actions need review.
- Plan
- Blocks file mutation and shell execution. Permitted research may contact external services; session state can still be saved.
- Work
- Edits files and runs commands within the permission you set.
- Operate
- Runs a fleet: several agents on one job, each in its role.
- Ask
- Prompts according to the active approval rules; saved permissions and hard policy boundaries still apply.
- Auto-Review
- Automatically reviews eligible actions and reports any action it cannot approve.
- Full Access
- Reduces approval prompts. It does not bypass hard policy boundaries or grant access outside the allowed scope.
Where to use Codewhale
The terminal and local browser client are available now. Desktop and hosted web apps are being developed around the same session model; their availability is separate. The terminal and local browser work without a Codewhale account. Hosted web and desktop access use an account, while the model connection remains your choice.
- Terminal
- ReleasedInstall GitHub release binaries for Linux, macOS, and Windows. npm and Cargo are alternatives. Android on Termux is a preview. The interactive TUI and codewhale exec for scripts ship together. Install guide
- Local browser
- Included with the terminalRun codewhale web to open Codewhale in your browser. Read conversations and tool results, send a task, and respond to approvals on your machine. Local browser guide
- Hosted web app
- Development previewSign in with a Codewhale account, then type /rc in a running terminal session to continue it from the web app. Hosted task execution is still being qualified. Sign in
- Desktop
- Development buildThe macOS app brings folders, conversations, and model connections into a desktop window. A public download is coming later.
- Cloud computers
- In developmentHosted computers for running your tasks.