# Nerve

A free, open-source desktop coding workbench for AI agents that keeps its own footprint small, running on an existing AI subscription or an API key.

Nerve is a desktop coding workbench for AI agents built to get more work out of every token: its generated base prompt is a 363-character block, only the tools a run can actually use are exposed to the model, tool results are capped before reaching the model, and long conversations compact automatically before the context window runs out.

Conversations, context, permissions, git operations, pull requests and background tasks all live in one window, visible and steerable while they run. Users sign in to a supported AI provider subscription over OAuth or paste an API key, can change the model, thinking level, agent mode or permission rule set mid-conversation, and can fork from any earlier point in a conversation without losing the original thread. Permission levels, read-only, supervised and autonomous, and rule sets are enforced as policy before a tool runs, not just as prompt instructions.

It runs on Linux, Windows 11 and macOS, is distributed under the Apache-2.0 license, and launches from npm with Node.js 24 or newer.

Website: https://nerve.tlmtech.dev/