gptme

gptme is the agent that goes where you already are. It runs on a laptop, over an ssh session, inside tmux, on a headless server and in a pipeline, and the trait that decides it daily is that it never assumes a graphical anything. One install and the same tool answers in all five places.

It handles one agent at a time, so it is a helper rather than a team, and the feature list is wide enough that you will not use half of it. Pick it when the terminal is your whole environment. Pick Aider when the work is a repository and you want the tighter tool.

The risk is that local-first means unprotected by default. Code executes in the environment you launched from, with no container in the capability list, so the same property that makes it useful over ssh makes a mistaken command a real event on a real host. On a shared server that is somebody else's problem too.

Run it as a user with the permissions the task needs and nothing more. What it does right: it commits automatically, so the state before a bad turn is recoverable with git rather than with memory, and pre-commit integration means the project's own checks run against what it wrote.

1. Context is gathered explicitly. Retrieval over local files is its own tool the model invokes, so what entered the window is visible in the transcript instead of assembled invisibly. 2. Changes are applied incrementally through a patch tool, with a separate faster path for bulk edits. 3. Guidance is separated from instructions through a lessons mechanism that matches on keywords, tools and patterns.

No benchmark accompanies any of this, and the release history instead documents dated capability additions from 2023 onward. For a tool of this size that is the appropriate evidence.

There is nothing to value here and that is the point. The first commit predates the category, the project has shipped continuously since, and it is one person's work with no entity attached and no monetisation attempted. Weekly package installs sit in the thousands against a modest star count, which is the ratio of a tool people actually use rather than one they bookmark.

Nobody acquires this and nobody shuts it down. The failure mode is a maintainer losing interest, not a board losing patience. Likely path: it keeps existing. Position: an unusually safe dependency precisely because there is no business model to change.

The interesting sentence in the description is that it runs in continuous integration. That moves it from a thing individuals install to a thing a pipeline calls, which is where measurable value lives. Cost for sixty people is model spend only, with no seats to negotiate and nothing to renew.

Against that: no identity system, no central log, no vendor to send a questionnaire to, and support is one maintainer with a chat channel. Onboarding is a single install command, so training cost is near zero. Approved with conditions: use it as a pipeline utility with a scoped token, not as the tool we standardise sixty desktops on.

MIT, and it was built by someone with my priorities. Models come from any of the hosted vendors, a router for a hundred more, or llama.cpp serving locally, so nothing about the loop requires an account. Extensions are plain Python packages providing tools, hooks and commands, protocol servers are discovered and loaded at runtime rather than declared up front, and a companion package adds tree-sitter code navigation as further servers.

There is even a variable that gives each tool call its own notification sound. Small, unnecessary, entirely in the spirit of a tool built to be lived in.