CodeLeveler

A local-first coding agent that works in your repository and leaves reviewable changes

LLM Mart 3 views 320k listing impressions

Give CodeLeveler a task to read code, edit files, and run tests. Inspect the changes as it works and resume your session later.

Use the terminal, browser, or command line with BigModel, DeepSeek, Moonshot/Kimi, OpenAI, Anthropic, or other OpenAI-compatible services. Sessions stay on your machine; model requests go to your configured provider.

Install

macOS and Linux:

curl -fsSL https://raw.githubusercontent.com/dengmengmian/CodeLeveler/main/install.sh | sh

The default install directory is ~/.local/bin; make sure it is on your PATH.

Or use Homebrew (macOS and Linux x86_64):

brew install dengmengmian/tap/leveler

On Windows, download the matching archive from Releases, verify its supplied SHA-256 checksum, extract it, and add the directory to PATH. Keep leveler.exe and leveler-confine.exe together.

On Linux, install bubblewrap before using restricted commands, for example with sudo apt install bubblewrap.

Get started

leveler login              # choose a provider and configure an API key
cd your-project
leveler

Start with a clean Git working tree so changes are easy to review. Run leveler doctor to diagnose configuration or environment issues.

Task Command
Open the terminal UI leveler
Open the browser UI leveler web
Run a task directly leveler run "find the failing tests and fix them"
Resume a session leveler resume
Show help leveler --help

leveler run --parallel 2 preserves the selected permission profile, network restriction and model-step ceiling, and supports --output jsonl. Parallel runs reject shipping flags (--commit, --push, --pr, --branch, --pr-base) before starting; ship the reviewed integration separately. For single-task runs, shipping works with either output format and shipping failures return a non-zero exit code. Candidate setup, execution or commit errors fail the run with their original cause before integration; the run still records candidate completion and cleans up its worktrees. Parallel candidates inherit the parent state home. Cleanup failures retain their paths in warnings and the persisted terminal record; they do not undo integration or replace the primary task error.

In the terminal, /goal <task> runs an explicit goal to completion, with the completion gate deciding when it is done. Use /help for commands and shortcuts, and /thinking to adjust reasoning intensity.

Multiple TUIs in the same project share its runtime. If an upgrade finds that runtime busy, a client with a compatible protocol and matching configuration opens against the existing runtime; a later startup retries the upgrade once it is idle. Normal TUI exit first cancels any remaining turn and stops background tasks in the current session, leaving other sessions running. Exit waits for cleanup to finish and reports failures instead of silently disconnecting. A lost connection or forcibly terminated client does not stop the shared runtime's tasks.

Tool execution summaries append elapsed time from actual admitted dispatch, update once per second, and freeze at the final execution boundary. Reconnects restore the runtime clock. Old history and detached process lifetimes without reliable execution timing keep their original summaries; unrelated clocks are never combined. Approval dialogs initially select Allow once; Esc still denies.

Permissions and privacy

The default Auto mode allows ordinary development actions and asks for approval for dangerous operations. Choose Restricted for stricter approval, or Full to allow all operations in an environment you trust.

The model may read workspace content and send relevant material to your provider. Keep secrets outside workspaces accessible to the agent. See the security policy for permission and platform limitations.

Update

For Homebrew installs:

From the project's README.

Comments (0)

Sign in to join the conversation.

No comments yet.