What is Nerve?
Nerve is an open-source desktop coding harness for people who want an agent’s work to stay visible and controllable. It combines a focused agent runtime with a workbench for conversations, files, context, Git, pull requests, background tasks, logs, settings, and project notes.
Nerve is not a code editor and does not replace your repository or IDE. It works directly with a directory you select, previews files in the workbench, and can open a project in VS Code or Zed.
The local-first shape
Section titled “The local-first shape”By default, the Electron app starts and supervises a daemon on 127.0.0.1:3747, then loads the bundled browser workbench. Nerve state lives under ~/.nerve; your source code remains in your project directory. The Electron browser profile stays in the operating system’s normal Electron profile location, outside NERVE_HOME.
The daemon reaches external services only when required by your selected model, a Git remote, or an integration or network tool you enable and invoke. Local-first does not mean offline-only.
The main concepts
Section titled “The main concepts”- Project: a Nerve record pointing to a local directory. Selecting another project is a workbench action; there is no server-wide current project.
- Conversation: persisted agent history associated with a project. Conversations can branch when you navigate to an earlier entry.
- Agent: the model, thinking level, mode, permission level, tools, and loaded resources used for work.
- Run: one durable execution attempt. Runs can stream output, pause for review, retry, or continue from a recoverable checkpoint.
- Task: a supervised long-running process such as a server or watcher. Tasks are distinct from finite shell commands and agent to-dos.
Why the controls matter
Section titled “Why the controls matter”Choose coding or planning mode and one of three authority levels:
- Read-only blocks mutation, command execution, ordinary network calls, and child-agent spawning.
- Supervised asks before non-read actions; selected read operations can be autoapproved.
- Autonomous runs allowed actions without approval.
These controls are enforced by Nerve’s tool policy, but they are not an operating-system sandbox. Review commands and integrations according to the trust you place in the model and project.