shim
Tiny local-first AI harness: any model, MCP and Agent Skills, natural-language CLI. Zero dependencies. MIT.
- Stars
- 434
- Forks
- 2
- + today
- +8
- Created
- 2d
Ranking data as of November 3, 2026 (UTC).
Star History
Today, hour by hour
Overview
Tiny local-first AI harness: any model, MCP and Agent Skills, natural-language CLI. Zero dependencies. MIT. It ranks #19 on GitTiger, gaining +8 stars on November 3, 2026 (UTC).
The project is written in Go and has 2 forks. It was created 2d ago.
What it is
Balanov/shim is a compact AI harness written in Go and released under the MIT licence. It is built to run locally and aimed at people who are happiest in a terminal. The repository description adds that it works with any model, supports MCP and Agent Skills, offers a natural-language CLI and has zero dependencies.
What the README preview describes
- **Two input styles:** experts get a complete scriptable CLI, while everyone else can just type or speak ordinary language. - **Command routing:** a small local model turns management requests into a command from an allowlist. Anything that mutates state asks for confirmation first. - **Model choice:** adapters for OpenAI-compatible and Anthropic APIs are built in, along with presets covering well-known hosted and local providers. - **Extensibility:** MCP servers can be attached over stdio or Streamable HTTP, and SKILL.md files are read on demand from common directories.
The preview is cut off after these items, so other features were not supplied.
Getting started
No install or run commands appear in the supplied data. Head to the repository README at https://github.com/Balanov/shim for setup.
git clone https://github.com/Balanov/shim.git
cd shim
# see README for setupFrequently Asked Questions
Which models can Shim work with?
Built-in adapters cover OpenAI-compatible and Anthropic APIs, and the preview also mentions ready-made presets for well-known providers, both hosted and local.
Does it support MCP?
Yes, per the README preview: MCP servers over stdio and Streamable HTTP, and Agent Skills loaded from SKILL.md files.
Can natural-language commands change things unsafely?
The README preview says a local model picks from an allowlisted set of commands and that mutating ones need confirmation. No further safety details were supplied.
What are the language and licence?
The metadata lists Go and MIT.