Nerve

A free, open-source coding workbench for AI agents — built to get more work out of every token.

Run it on the AI subscription you already pay for, or your own API key. The harness keeps its own footprint small: a 363-character generated base prompt, only the tools a run can actually use, tool results capped before they reach the model, and automatic compaction before the window runs out.

Everything else — conversations, context, permissions, Git, pull requests, and background tasks — lives in one window, visible and steerable while it runs, and stays on your disk.

Website · Documentation · Install guide · Contributing · Security

[!NOTE] Nerve is beta software. It runs on Linux, Windows 11, and macOS and is distributed under the Apache-2.0 license.

Quick start

Nerve requires Node.js 24 or newer. Launch the latest published desktop app directly from npm:

npx @nervekit/desktop@latest
# or
pnpm dlx @nervekit/desktop@latest

The first launch may download Electron's platform binary. Nerve starts a local loopback daemon by default, and its application data stays under ~/.nerve.

Highlights

• A lean harness. The generated base prompt is one 363-character block; tool rules, planning instructions, and project context are appended only when they apply, and gated tools never reach the request at all.
• Bounded context, not a dead end. Tool results reaching the model are capped at 24,000 bytes while the complete output stays retrievable on disk, and long conversations compact automatically instead of ending.
• Use the AI access you already have. Sign in to a supported provider subscription over OAuth, or paste an API key. Credentials are stored encrypted on your machine and the model catalog comes from the provider.
• Steerable, branchable conversations. Change the model, thinking level, agent mode, or permission rule set mid-conversation — it applies from the next provider request — and fork from any earlier entry without losing the original lineage.
• Permissions enforced in policy, not prompts. Read-only, supervised, and autonomous levels, planning mode, rule sets, and per-project or per-conversation overlays, evaluated before a tool runs.
• Multiple projects and repositories. Repositories are discovered beneath a workspace root, each with its own branch and working-tree state, Git operations, and GitHub pull requests.
• Delegation with oversight. Supervised background tasks with retained logs, and bounded read-only Explore sub-agents that research in parallel and report back.

Permission policy is Nerve's tool authorization and review layer, not an OS or container sandbox. Local-first is not offline-only: provider, Git, voice, web, and integration calls leave your machine when you use them.

Documentation

The website is the primary source for product and developer documentation:

• Get started
• Use the workbench
• Configure and operate Nerve
• Troubleshoot installation and runtime issues
• Understand the architecture
• Read the Protocol v1 reference

Develop from source

The repository requires Node.js 24 or newer, pnpm 11.20.0, and rustup; the Rust version is pinned in rust-toolchain.toml.

pnpm install
pnpm desktop

Use pnpm dev for the daemon and browser UI development servers. See the development guide and CONTRIBUTING.md for the complete workflow. Release engineering details remain in docs/runbooks/release.md.

Support

If Nerve is useful to you, starring the repository is the cheapest way to help other developers find it. You can also support its continued development on Patreon.

Contributing, security, and license

Contributions are welcome. Read CONTRIBUTING.md before opening a change, and report vulnerabilities through the private channels in SECURITY.md.

Nerve is licensed under Apache-2.0. See LICENSE and NOTICE.

Acknowledgements

Nerve's model routing, provider integrations, and streaming are built on @earendil-works/pi-ai, a unified LLM API client by Mario Zechner (MIT license).