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, editable project files, context, Git, pull requests, profile-based sub-agents, launch configurations, asynchronous Bash, 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: a model-driven worker with a thinking level, mode, permission level, tools, and loaded resources. Each conversation has a lead; it can delegate to sub-agents, each created from a profile such as Explore (read-only research) or Developer.
- Execution: a period of model-driven work in a conversation. It can stream output, pause for review, retry, or be interrupted; stored messages and tool results remain available afterward.
- Launch configuration: a saved command, working directory, optional port, and run policy for a server or watcher. Each start creates a separately inspectable launch run.
- Async Bash: a conversation-owned command that returns immediately or moves to the background after a configured wait. It reports completion without polling.
- Profile: a creation-time template for a sub-agent, not a live binding that reconfigures existing children.
Why the controls matter
Section titled “Why the controls matter”Choose coding or planning mode and one of three authority levels:
- Read-only allows local inspection and interaction, and explicitly permits enabled sub-agent prompting and stopping. A child’s profile determines its own authority; disable autonomous profiles for inspection-only delegation.
- Supervised automatically allows safe reads and asks before other actions unless a matching permission rule allows the action.
- Autonomous runs allowed actions without approval, while explicit block exceptions still apply.
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.