gptme is an open-source, terminal-first AI agent designed for software engineering and system tasks, originally created by Erik Bjäreholt. Operating directly in standard terminal sessions, remote SSH environments, tmux sessions, and CI/CD pipelines, it acts as an unconstrained assistant capable of running shell commands, executing Python scripts, modifying local repositories, and automating browser tasks. In standard workflows, developers interact with gptme via an interactive terminal REPL, single-shot CLI commands, or a self-hosted web UI. The agent is provider-agnostic, supporting models from Anthropic (such as Claude 3.7 Sonnet), OpenAI, Google (Gemini), xAI (Grok), DeepSeek (R1), and OpenRouter, as well as offline local models via llama.cpp. When given a task, gptme inspects directory structures, generates targeted code edits using incremental patch and morph tools, runs test suites directly in your environment, and reads execution output to self-correct failures. What differentiates gptme is its extensive system integration and layered extensibility. It features an automated "lessons" engine that dynamically injects relevant context and rules into the prompt based on tool usage and keyword patterns, native support for Model Context Protocol (MCP) tools, background job execution, and tree-sitter-based structural code analysis (gptme-codegraph) for call graph and impact analysis. It also supports computer vision, desktop automation, and subagent orchestration for delegating complex workflows. gptme is free, open-source software distributed primarily through PyPI with standalone desktop packages available across Linux, macOS, and Windows. Users can connect standard provider API keys or use web-subscription auth without vendor lock-in.
Tags: ai agent, cli, coding agent, OpenSource, terminal