Codewhale Terminal Coding Agent
Delegate coding tasks in the terminal so an agent can inspect a repository, edit files, and run commands.
- Source repo
- codewhale-hq/Codewhale
- Stars
- ★ 41k
- Last updated
- today
- License
- MIT
- Primary language
- Rust
- FA score
- 75/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude APIChatGPT · Claude Code (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer who wants help fixing failing tests and an explanation of the changes can assign the task from the repository root and inspect the diff.
- Not a fit if
- Teams that require an OS sandbox on Windows
- Users who need a native desktop app today
- Source review
- 75/100 · Some gaps
What does this agent do, and when should you use it?
Codewhale is an open-source coding agent built in Rust, primarily used through an interactive terminal app, with a one-task CLI, Web UI, Runtime API, and ACP interface also available. It reads and searches files in the current repository, edits files and runs shell commands according to the selected permissions, and continues until the task is done or it needs input. It has more than 40 built-in provider integrations, supports OpenAI-compatible endpoints, and can use local models through Ollama, vLLM, or SGLang. /receipts reports file changes, commands, web and MCP calls, approvals, and failures; tool-enabled runs can also be inspected after exit. The software is MIT-licensed and does not require an account, but hosted models can incur provider charges; Windows does not currently have an OS-level sandbox.
Start codewhale at a Git repository root and give it a task; it reads and searches repository files, then edits files and runs commands in Work or Operate mode. Under Ask permission it asks before running non-read-only shell commands; Auto-Review and Full Access are also available. Select a model with /provider and /model, or connect a local Ollama, vLLM, or SGLang runtime. Tools can be added through MCP servers, plugins, skills, and hooks. Other interfaces include codewhale exec for non-interactive tasks, codewhale review --pr N for advisory GitHub pull request review, codewhale web for a local Web UI, the Runtime API, and ACP. Outputs include file changes and records of commands, tool calls, approvals, and failures, available through /receipts or codewhale receipts --last.
- A developer who wants help fixing failing tests and an explanation of the changes can assign the task from the repository root and inspect the diff.
- An engineer who needs to explore and plan without changing files can use Plan mode to read and search the code.
- A maintainer automating a one-off coding task in a script or CI can run
codewhale execwith chosen access and sandbox options. - A team reviewing GitHub pull requests can run
codewhale review --pr Nfor advisory feedback. - A developer connecting team MCP tools or parallel agents can configure MCP servers, plugins, skills, workflows, or sub-agents.
How do you install or deploy this agent?
On macOS and Linux, install the latest release with the installer script; it places codewhale and codew in ~/.local/bin. Windows installation options listed in the README include release distribution channels; npm and Cargo are also supported.
How do you use this agent?
From the root of a Git repository, start the terminal app:
codewhaleUse /provider in the interface to choose a provider, add an API key, and select a model. For local models, use Ollama, vLLM, or SGLang. Then enter a task, such as fixing failing tests. With the default Ask permission, file edits are applied and non-read-only shell commands require approval.
For a script or CI, run:
codewhale exec --auto --sandbox workspace-write "fix the failing tests and explain what changed"What are this agent's strengths and limitations?
- More than 40 built-in providers, OpenAI-compatible endpoints, and local models through Ollama, vLLM, or SGLang give teams choices across model and deployment setups.
- The interactive terminal,
exec, Web UI, Runtime API, and ACP share one engine, supporting both hands-on and automated use. /receiptstracks changed files, commands, web and MCP calls, agents, approvals, and failures for task review.- Plan, Work, and Operate modes plus Ask, Auto-Review, and Full Access permissions offer distinct ways to control execution.
- Hosted models require a provider key and may incur provider charges; local models require the user to run a supported inference runtime.
- Windows has no OS-level shell sandbox yet, and Full Access disables Codewhale's sandbox.
- Telemetry is on by default; the README says it excludes conversations, code, prompts, and credentials, but users must turn it off to disable it.
- The native desktop app is in development and has no public download.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Codewhale Terminal Coding Agent This agent | 75 · Some gaps | CLIFree + model costs | ★ 41k | today | Rust | OpenAI API · Claude API |
| Empryo | 50 · Major gaps | CLIFree + model costs | ★ 1.4k | 5d ago | TypeScript | OpenAI API · Claude API |
| Aster | 77 · Good | CLIFree + model costs | ★ 131 | 2d ago | Rust | OpenAI API |
| VT Code | 66 · Some gaps | CLIFree + model costs | ★ 883 | today | Rust | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows configurable tool permissions: Ask confirms non-read-only shell commands, while Full Access bypasses the sandbox. The sandbox document describes platform-specific selection and fallback; receipts and undo are documented. Telemetry fields and opt-out, local key storage, and request preview are disclosed. Deductions: file edits proceed without confirmation by default, so users must understand the mode differences; dependency materials give versions and locked-install guidance but no vulnerability audit or remediation evidence; publisher identity is unknown and maintenance attribution is limited.
The README, sandbox guide, and install checklist are broadly consistent and distinguish wired behavior from future work; diagnostics, probe failures, and platform fallbacks are described. Deductions: these are static descriptions, not evidence validating runtime outcomes; external services and OS sandbox availability remain environment-dependent.
The repository addresses terminal developers, scripts and CI, classroom deployment, and editor integrations, with multiple model providers, modes, and extensions. Platform sandbox boundaries, plugin limits, and unimplemented workroom features are explicit. Deductions: examples cover common triggers, but guidance for choosing among complex configuration and extension combinations is incomplete; Windows currently has no command OS sandbox.
Documentation is organized across installation, modes, providers, security, extensions, and APIs, with command examples, troubleshooting, limitations, an MIT license, and a changelog entry point. Deductions: readers must track multiple install channels and distinguish main from released versions; compatibility and maintenance notes exist, but unreleased branches receive only best-effort support and publisher identity is unverified.
The product can read and edit repositories, invoke tools, and provide receipts, diffs, undo, multiple providers, and a script interface, producing outputs useful for review and automation. Deductions: value depends on model/API costs, platform sandbox availability, and operator configuration; high-privilege automation carries costs and risks the operator must manage.
Many claims point to concrete settings, commands, source paths, statuses, and limitations, and the materials distinguish implemented capabilities from future work. Deductions: although several file types are provided, most product and security claims in this static review are repository documentation statements without independent implementation evidence in the supplied material.
- Full Access and `exec --auto` can run commands without per-command confirmation or the Codewhale command sandbox; check the permission mode and actual sandbox availability first.
- Telemetry is enabled by default; disable it if you do not want anonymous usage counts. Data handling and isolation with external model providers or OpenSandbox also depend on those services.
- The documentation says Windows currently has no Codewhale OS command sandbox; approval policy is not operating-system isolation.
FAQ
Does Codewhale require payment or an account?
Can it run fully offline?
Will it run commands without asking me?
exec --auto defaults to Full Access, and --sandbox workspace-write restricts execution.