原文: Your agent in your terminal, equipped with local tools: writes code, uses the terminal, browses the web. Make your own persistent autonomous agent on top!
📜 A personal AI agent that runs anywhere a terminal runs — your laptop,
ssh sessions, tmux, headless servers, CI pipelines.
Provider-agnostic, local-first, and unconstrained: ships with shell, Python, web,
vision, and everything else an agent needs.
A great coding agent, but general-purpose enough to assist in all kinds of knowledge-work.
Free and open-source. Works with Anthropic, OpenAI, Google, xAI, DeepSeek, OpenRouter,
or fully local via llama.cpp — your data, your models, your terminal.
A capable alternative to Claude Code,
Codex, Cursor, and Warp — one of the first agent CLIs (Spring 2023), still in very
active development.
The screencasts below are from 2023. gptme has evolved a lot since then!
For up-to-date examples and screenshots, see the Documentation.
We're working on automated demo generation: #1554.
Fibonacci
Snake with curses
Steps
Create a new dir 'gptme-test-fib' and git init
Write a fib function to fib.py, commit
Create a public repo and push to GitHub
Steps
Create a snake game with curses to snake.py
Running fails, ask gptme to fix a bug
Game runs
Ask gptme to add color
Minor struggles
Finished game with green snake and red apple pie!
Mandelbrot with curses
Answer question from URL
Steps
Render mandelbrot with curses to mandelbrot_curses.py
Program runs
Add color
Steps
Ask who the CEO of Superuser Labs is, passing website URL
Skills — lightweight workflow bundles (Anthropic format) that auto-load when mentioned by name. Great for packaging reusable instructions and helper scripts without writing Python.
Lessons — contextual guidance that auto-injects into conversations based on keywords, tools, and patterns. Write your own to capture team best-practices or domain knowledge.
Hooks — run custom code at key lifecycle events (before/after tool calls, on conversation start, etc.) without a full plugin.
gptme-contrib — community-contributed plugins, packages, scripts, and lessons:
pipx install gptme # MCP support included by default
gptme can discover and dynamically load MCP servers, giving the agent access to databases, APIs, file systems, and any other MCP-compatible tool. See the MCP docs for server configuration.
This makes gptme available as a drop-in coding agent in Zed and JetBrains IDEs. Your editor sends requests, gptme executes with its full toolset (shell, browser, files, etc.) and streams results back.
🤖 Autonomous Agents
gptme is designed to run not just interactively but as a persistent autonomous agent — an AI that runs continuously, remembers everything, and gets better over time. The gptme-agent-template provides a complete scaffold:
Persistent workspace — git-tracked "brain" with journal, tasks, knowledge base, and lessons
Run loops — scheduled (systemd/launchd) or event-driven autonomous operation
Task management — structured task queue with YAML metadata and GTD-style workflows
Meta-learning — lessons system captures behavioral patterns and improves over time
Multi-agent coordination — file leases, message bus, and work claiming for concurrent agents
External integrations — GitHub, email, Discord, Twitter, RSS, and more
# Create and run your own agent
gptme-agent create ~/my-agent --name MyAgent
gptme-agent install # runs on a schedule
gptme-agent status # check on it
Bob is the reference implementation — a production autonomous agent that's been running continuously since late 2024. Bob opens PRs, reviews code, fixes CI, manages his own task queue, maintains a growing set of behavioral lessons, posts on Twitter, responds on Discord, and writes blog posts.
Multiple specialized agents can run in parallel — e.g. Bob (engineering) and Alice (personal assistant & orchestration) — coordinating through shared infrastructure.
Persistent agents need guardrails around the full loop, not just tool permissions:
Input guardrails — structured task selectors in the agent workspace keep work focused and reduce thrashing on notifications or ambiguous work. Bob uses a CASCADE-style selector for this layer.
Pre-action guardrails — lessons inject situational guidance before the agent acts.
Output guardrails — hooks and pre-commit checks validate file changes before control returns to the user.
This stack is simple and composable: selectors improve work choice, lessons steer behavior, and checks verify the result. You can add evals on top later, but the baseline guardrail loop already exists.
🛠 Use Cases
🖥 Development: Write and run code faster with AI assistance.
🎯 Shell Expert: Get the right command using natural language (no more memorizing flags!).
📊 Data Analysis: Process and analyze data directly in your terminal.
🎓 Interactive Learning: Experiment with new technologies or codebases hands-on.
🤖 Agents & Tools: Build long-running autonomous agents for real work.
