agentskillexchange gptme — Terminal AI Agent with Local Tools
gptme is an open-source framework for building and running LLM-powered agents directly in your terminal. It equips agents with real local tools — shell execution, file editing, browser control, Python REPL, and persistent context — and supports multiple model providers out of the box.
Unlike browser-based AI assistants, gptme agents run natively in your environment and can read your files, run commands, browse the web, and persist state across sessions. It is designed for autonomous operation: agents can be left running, pick up work from structured task files, and chain tool calls without per-step human confirmation.
Key Features
- Multi-provider: OpenAI, Anthropic Claude, Google Gemini, OpenRouter, Ollama (local), llama.cpp, and any OpenAI-compatible endpoint
- Rich tool set: Shell (
bash), file read/write/patch, Python REPL, browser automation (Playwright),computer(screenshot + click), RAG over local documents - Persistent sessions: Conversations are stored locally; resume any session by name
- Autonomous loops: Built-in support for long-running unattended agents with structured task management and journaling
- Local-first: Works fully offline with local models; no cloud dependency required
- Extensible: MCP server support, custom tools, skill injection via
gptme.toml
Installation
Install or set up from the source-backed instructions:
Install with uv tool install gptme.
- Source: https://github.com/gptme/gptme
Basic Usage
Because gptme can run shell commands, edit files, control a browser, and persist sessions locally, run it with an explicit operating boundary. Prefer a project sandbox or worktree, review enabled tool permissions before unattended runs, and avoid pointing agents at sensitive directories unless you intend those files to be inspected or modified.
# Start an interactive session
gptme "Write a Python script to list all .md files in the current dir"
# Run non-interactively (for scripting and automation)
gptme --non-interactive "Fix the failing test in tests/test_api.py"
# Use a specific provider / model
gptme -m anthropic/claude-opus-4-5 "Review the PR diff and summarize findings"
# Use a local model via Ollama
OPENAI_API_BASE="http://localhost:11434/v1" gptme -m local/llama3.2 "Hello"
# Resume a named session
gptme --name my-project "Continue where we left off"
Autonomous Agent Setup
gptme ships an agent template for building persistent autonomous agents with task queues, journals, and structured work loops. See gptme-agent-template for the scaffold.
# Install gptme-agent-template to bootstrap a new agent workspace
git clone https://github.com/gptme/gptme-agent-template.git my-agent
cd my-agent
./scripts/fork.sh . MyAgent
gptme "Hello, I am MyAgent. What shall we build today?"