Let an AI agent automatically refactor code, run tests, and commit changes in your project folder with your approval.
Have the agent research a feature request, propose a structured implementation plan, and review the code before you merge.
Chain a planner, executor, and reviewer together so the agent researches, implements, and validates changes in one workflow.
Track tasks, permissions, and conversation history for each project in a local database without sending data to the cloud.
Requires Node.js 22+ and npm 10+. You'll need an API key from OpenAI, Anthropic, or another supported model provider to run agents.
Little Peanut is a desktop AI workbench built as an Electron application with a React interface and a local SQLite database. Its goal is not to be another chat client. The README frames it as a place where a model can actually do work inside a real project folder on your computer, rather than just answering questions in a browser tab. The repo is described as early development, with the basic skeleton in place and many features still being filled in. The app organises work into named modes. Chat is for light conversation, Agent runs tasks against a project, Plan is a read-only mode that does research and proposes a structured implementation plan without changing files, and Pipeline strings together a planner, executor, and reviewer in stages. There are also modes called Review, Writing, Code, Learning, Research, Brainstorm, Translate, and Summarize, each tuned for a different kind of request. The agent can call local tools such as file read and search, file edits, terminal commands, todo and task tracking, web search and fetch, lint reading, and a memory store. Tool calls go through a permission gate. Risky or destructive actions trigger an approval prompt, and permission rules can be saved per project or per session. The README is careful to say this is not a full sandbox, only a controllable boundary on a local machine. Little Peanut supports several model providers including OpenAI, Anthropic, Gemini, and OpenAI-compatible endpoints, with custom base URLs and API keys. It can also connect to MCP servers over stdio, SSE, or HTTP for extra tools. SQLite stores conversations, messages, tool runs, tasks, permission rules, a memory index, cost logs, model and provider config, and channel bot settings for QQ, Feishu, and DingTalk integrations. To run it you need Node.js 22 or newer and npm 10 or newer. The root folder is a small delegator, and the real Electron app lives in the src directory. Typical commands are npm install, npm run install:app, npm run dev to start, and npm run build or npm run dist to package.
Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.