Warning

This project gives an agent the ability to run shell commands and modify files on your system. You should not use it unless you understand the behaviors and precautions taken by the mcp-edit and mcp-shell crates.

Also... be aware that a large amount of code in this project was generated.

Quick Start

Install with cargo

> cargo install --git https://github.com/dstoc/llment llment mcp-edit mcp-shell

Run with defaults (preferred):

• assumes you have llama-server on localhost:8000 serving gpt-oss (without the --jinja flag)
• no external MCP servers, only the harmless builtin tool get_message_count

> llment

Providers

--provider can be used to select a different provider:

• harmony (preferred) connects with openai/harmony compatible models via the llama-server /completion API
• ollama - uses ollama-rs to interface with the Ollama API
  • Example

    > llment --host http://localhost:11434 --provider ollama --model qwen3:latest

• openai-chat - uses async-openai to connect with OpenAI's /v1/chat/completions API
  • Example - Ollama via OpenAI compat API

    > llment --host http://localhost:11434 --provider openai-chat --model qwen3:latest

• gemini-rust - uses the gemini-rust crate to interface with the Gemini API

  • Status: functional but incomplete. Reasoning context is not maintained.

  • Requires GEMINI_API_KEY in env.

  • Example

    > GEMINI_API_KEY=... llment --provider gemini-rust --model gemini-2.5-flash

--model and --host can be used to customize further, e.g.

Ollama (with custom model and endpoint):

> llment --provider ollama --model qwen3:30b --host https://my-ollama.tailc.ts.net:11434

Model Context Protocol servers

Warning

There are currently no approval steps in order for an agent to execute functions exposed by MCP servers.

--mcp file.json loads a claude-code like mcp.json file.

For example, the following configuration loads two STDIO based MCP servers. The functions from mcp-edit will be prefixed with files_ as in files_create_file, similarly with shell_ for mcp-shell. MCP server names must not contain underscores because tool identifiers are generated as <prefix>_<tool>. The server commands are launched with the same working directory that llment was.

{
  "mcpServers": {
    "files": {
      "command": "mcp-edit"
    },
    "shell": {
      "command": "mcp-shell",
      "args": [
        "--workdir", "/home/user/workspace",
        "--container", "sandbox"
      ]
    }
  }
}

No other "mcp.json" options or features beyond those used above are currently supported.

mcp-edit

The mcp-edit server provides a set of file system tools similar to gemini-cli.

Warning

By default, mcp-edit provides read-only access to the current working directory.

Usage: mcp-edit [OPTIONS] [WORKSPACE_ROOT] [MOUNT_POINT]

Arguments:
  [WORKSPACE_ROOT]  Workspace root directory (default: current directory) [default: .]
  [MOUNT_POINT]     Mount point path used in responses (default: `/home/user/workspace`) [default: /home/user/workspace]

Options:
      --trace               Show trace
      --allow-modification  Enable file modification tools
  -h, --help                Print help

mcp-shell

The mcp-shell server provides the ability to execute shell commands inside a container.

An example container configuration is provided in ./sandbox.

Usage: mcp-shell [OPTIONS] --workdir <WORKDIR> <--container <CONTAINER>|--unsafe-local-access>

Options:
      --container <CONTAINER>  Run commands inside a Podman container
      --unsafe-local-access    Run commands inside a local shell. Unsafe
      --workdir <WORKDIR>      Working directory for command execution
      --trace                  Show trace
  -h, --help                   Print help

Commands are executed in the named container using podman exec