geoffrey-llm 0.1.0

A lightweight toolkit for LLM fine-tuning, inference optimization, and agent development

geoffrey-llm

A lightweight toolkit for LLM fine-tuning, inference optimization, and agent development.

Features

• geocode - Claude Code-like coding assistant with interactive REPL
  • Multi-model support (Kimi, DeepSeek, Qwen, OpenAI-compatible)
  • Memory system with file-based storage
  • MCP (Model Context Protocol) integration
  • Tool calling (File read/write/edit, Bash)
• More features coming soon (LoRA fine-tuning, GraphRAG, etc.)

Installation

pip install geoffrey-llm

Install with all features

pip install geoffrey-llm[all]

Install geocode only

pip install geoffrey-llm[geocode]

Quick Start

Using geocode (REPL)

# After installing geocode
geocode

# Or specify provider
geocode --provider deepseek --model deepseek-chat

Python API

from geoffrey_llm.geocode import REPL

# Create REPL with default settings
repl = REPL()

# Or with specific provider
from geoffrey_llm.geocode.models.base import get_registry, ModelConfig

config = ModelConfig(model_name="moonshot-v1-8k")
registry = get_registry()
model = registry.create("kimi", config)
repl = REPL(model=model)

geocode Module

geocode is a Claude Code-like coding assistant with:

Features

• Interactive REPL - Terminal-based chat interface
• Multi-Model Support - Works with Kimi, DeepSeek, Qwen, and any OpenAI-compatible API
• Memory System - Persistent memory with file-based storage
• MCP Integration - Connect to MCP servers for extended capabilities
• Tool Calling - Built-in tools for file operations and shell commands

Supported Models

Provider	Model Names	Environment Variable
Kimi (Moonshot)	moonshot-v1-8k, moonshot-v1-32k, moonshot-v1-128k	KIMI_API_KEY
DeepSeek	deepseek-chat, deepseek-coder	DEEPSEEK_API_KEY
Qwen (DashScope)	qwen-turbo, qwen-plus, qwen-max	DASHSCOPE_API_KEY
OpenAI-compatible	gpt-3.5-turbo, gpt-4, llama3, etc.	OPENAI_API_KEY

Configuration

Create ~/.geoffrey/config.yaml:

model:
  provider: kimi
  model_name: moonshot-v1-8k
  api_key: ${KIMI_API_KEY}  # Or set env var directly

Commands in REPL

Command	Description
exit, quit	End session
/new	Start new session
/resume <id>	Resume existing session
/sessions	List all sessions
/memory save <type> <content>	Save a memory
/memory list [type]	List memories
/memory recall <query>	Search memories
/mcp list	List MCP servers

Memory Types

Type	Description
user	User preferences, identity
feedback	User corrections, feedback
project	Project-specific context
reference	External reference material

MCP Servers

Configure MCP servers in ~/.geoffrey/mcp.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
      "env": {}
    }
  }
}

Tool Calling

geocode supports tool calling with built-in tools:

• FileRead - Read files with line offset/limit
• FileWrite - Write content to files
• FileEdit - Edit files using search/replace
• Bash - Execute shell commands (sandboxed)

Memory System

Memories are stored as markdown files with YAML frontmatter in ~/.geoffrey/memory/:

---
id: mem_abc123
type: user
created_at: 2026-04-18T10:00:00Z
updated_at: 2026-04-18T10:00:00Z
tags:
  - preference
---
User prefers dark mode interface.

Session Management

Sessions are stored as JSON files in ~/.geoffrey/sessions/. Each session contains:

• Conversation history
• Project context
• Model/ provider settings
• Last active timestamp

License

MIT License

Roadmap

• LoRA/QLoRA fine-tuning interface
• Unified inference backend (vLLM, llama.cpp, TensorRT)
• GraphRAG-lite for knowledge graph RAG
• Log analysis Agent toolkit
• Model quantization utilities (AWQ, GPTQ)