Open source · MIT · Any model
An AI agent that lives in your terminal.
One agent you run yourself, extend with your own tools, and can leave running on its own. Anywhere a terminal runs — laptop, ssh, tmux, headless servers, CI — on any model, including local ones.
See it work
Your machine, your tools, no sandbox in the way.
It works in your real environment: your shell and Python, your files, your git and gh setup, a browser and the whole desktop when a task needs one. Point it at a local model and nothing leaves the machine at all.
pytest tests/test_users.py -x -qF 1 failed, 11 passed
- initial = name.strip()[0].upper() + name = name.strip() + if not name: + raise ValidationError("name is required")ValueError("empty") + initial = name[0].upper()
pytest tests/test_users.py -q12 passed
What it can actually do
Full tool reference →- shell
A persistent session: builds, tests, git, ssh, package managers — with state kept between commands.
- ipython
A stateful Python REPL with your installed libraries, so data stays loaded between steps.
- patch · morph
Surgical edits to existing files instead of rewriting them wholesale.
- browser
A real Chromium: search, read pages, click, fill forms, scroll, screenshot, read PDFs.
- computer
The desktop itself on X11 or macOS — mouse, keyboard and screen capture, for apps with no API.
- subagent
Spawns child agents in parallel or as a pipeline, each with its own context, and steers or cancels them mid-run.
- tmux
Long-running processes it can come back to: dev servers, log tails, REPLs.
- vision · screenshot
Reads images, diagrams, its own screenshots and rendered pages.
Beyond the prompt
Most of what gptme does, nobody is watching.
gptme is built to keep working when you close the laptop. One command turns it into a persistent agent with a memory, a task queue and a schedule — the same scaffolding the agents below run on.
- A workspace it owns
- Git-tracked journal, tasks, knowledge and lessons. Its memory is a repo you can read.
- Run loops
- systemd or launchd on a schedule, or triggered by events like a new issue or a failing build.
- It learns in writing
- Lessons it writes for itself get matched back into later sessions, so mistakes stick.
- More than one
- File leases, a message bus and work claims let several agents share a workspace without collisions.
On a headless box, gptme service init scaffolds the same thing as a systemd service. See running agents autonomously.
The reference agent
Bob has been running on gptme since 2024, without taking a day off.
He picks up his own work, opens and reviews pull requests, fixes CI, writes blog posts, and keeps the lessons that make his next run better. A sibling agent, Alice, runs the same template as a personal assistant. Fork either shape in a single command.
See Bob's timeline →- merged public PRs
- 4,000+
- since late 2024
- 24/7
Run it your way
Same agent and config on every surface
- $ gptme
CLI
Interactive, or non-interactive for scripts and CI.
- $ gptme-tui
Terminal UI
A full-screen interface for long sessions.
- chat.gptme.org
Web UI
Bundled with gptme-server. Point it at your own machine.
- linux · macos · windows
Desktop app
Native app with auto-updates.
- zed · jetbrains
Your editor
A drop-in coding agent over ACP.
- rest · mcp · acp
Server
A REST API, plus MCP and ACP in both directions.
Bring your own model
Never locked to one provider, or one price.
One flag switches model, provider or price tier — the tools, config and workflow stay identical. Already paying for ChatGPT Plus/Pro or SuperGrok? Use that plan instead of an API key, at no cost per token. Running something high-volume? Open-weight models like DeepSeek V4.1 Flash and GLM 5.3 Flash hold up in agentic work for a fraction of frontier prices, and one OpenRouter key reaches 100+ models with data collection denied by default.
Fits the stack you already have
Open protocols, no lock-in, MIT
- github · email · chat · voice
Reachable where you already are
Mention @gptme on an issue or pull request. Let an agent read and answer email, sit in Discord or WhatsApp, or take a phone call. Agents can also message each other over SSH.
- mcp · acp
MCP in both directions
Load any MCP server as gptme tools — or run gptme as an MCP server, handing its persistent shell, Python session and file tools to Claude Desktop, Cursor or anything else that speaks MCP.
- plugins · skills · lessons
Extend it without forking it
Plugins add tools, hooks and commands. Skills and lessons teach it your workflows in plain Markdown. gptme-contrib carries community packages: code graphs, multi-model consensus, LSP, image generation.
- GitHub stars
- 4.4k
- Contributors
- 50
- License
- MIT
- First commit
- 2023
Stars and contributors update daily from gptme/stats.
Run it yourself
Your machine, your keys, your data.
One pipx install gptme and it is yours: MIT-licensed, no account needed, and a local model if you want nothing leaving the machine at all.
Or let us run it
gptme.ai runs the same agent, hosted.
Sign in and go: no install, no provider keys of your own, GitHub connected, same tools. Or keep running gptme locally and use it only for model access. In early access.
Go to gptme.ai →