Image: GitHub
Why it mattersAnyone who wants a local agent on a laptop can try one that is 7 MB. A heavy framework install is not the only way in.
A one-person agent harness written in Go reached 189 GitHub stars six days after it was pushed to GitHub. The project is called Shim, by GitHub user Balanov. The licence is MIT, and the repository has twelve forks. The whole thing is one static binary with zero dependencies, the README states.
What it is
Shim is a local-first AI harness, the README says. It runs in a terminal and talks to a model over HTTP. A natural-language mode lets a user type what they want. A full CLI mode runs a scriptable command for the same job. A script can call the tool without a model in the middle. A small local model, chosen by the user, picks which allowlisted command fits the request.
The project calls out ten model providers by name: Ollama, LM Studio, llama.cpp, OpenAI, OpenRouter, Anthropic, Gemini, GitHub Models and Hugging Face. Any other OpenAI-compatible endpoint also works, the README says. One tool, one config file, every model behind one command.
What it reports about itself
The README publishes two measured numbers for the v0.1.0 release. The static binary is 7.29 MiB after stripping. The startup time is 9.96 ms, taken as the median of 30 runs of shim --version on Windows. Both numbers are for the author's own machine and nobody else has reproduced them. A reader who tests on a different machine should expect a different number. The README does not state the hardware.
How it runs
Shim is a single process. There is no background daemon and no TUI, the README states. State lives in two local files: a plain JSON config and a JSONL session log. A user can read either in any text editor. Nothing calls home. No telemetry.
Permissions are explicit. A read-only mode lets the tool read files but not write. An ask mode pauses before every write. An auto mode acts without stopping. Model Context Protocol servers connect over stdio or HTTP. The same MCP server a user already runs for Claude Code or Cursor works here. Agent Skills live in SKILL.md files that load on demand.
Where it fits
The project is a minimal alternative to the agent harnesses already in the market. The README does not name the competition. It supports the same model list and speaks MCP. It runs on the same Windows, macOS and Linux machines that bigger tools already target. The difference is the size. 7.29 MiB and one binary is a smaller surface than any Python framework a user already knows.
The repository was created on 2 November and has one commit. The maintainer has three open issues. Go 1.25 is the install requirement for building from source.
Source
Balanov/shim, GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
