Free AI Programming Assistant (Golang)

简体中文

Agent-code-assistant is a code-assistant (LLM + Tool + Loop)

A minimalist, auditable, and privatizable programming assistant CLI implemented using the Go standard library, without relying on any third-party Golang frameworks, packages, or Agent SDKs.

⚠️ **Because this involves reading and writing local files, please verify the code before running it to ensure security and avoid any loss of local files! **

---

Golang version

Golang： go 1.25.5

✨ Features

• ✅ Does not use any third-party Go libraries

• ✅ Configuration file driven (config.yaml)

• ✅ Supports tool calls (file read/write/directory listing)

• ✅ Can connect to large local or remote models (Ollama/DeepSeek/OpenAI style)

• ✅ Supports user input length limits (prevents excessively long input)

• ✅ Simple code structure, fully understandable, and rewriteable at any time

---

The assistant is essentially just a loop

The core process consists of only these 5 steps:

↓ LLM generates a response (may include tool calls)

↓ Execute the tool locally

↓ Feed back the tool results to the LLM

↓ Output the final answer

No Agent magic, no framework black box.

---

📁 Project Structure

.
├── main.go       # Main program (single file)
├── config.yaml   # Model and behavior configuration
├── main_test.go  # Unit testing
└── README.md

---

⚙️ Configuration Instructions (config.yaml)

model: deepseek-r1
endpoint: http://127.0.0.1:11434/api/generate
temperature: 0.2
system_prompt: You are a free programming assistant.
input_limit: 0

Parameter Description

Parameter	Description
model	Model name
endpoint	Model HTTP interface address
temperature	Generated temperature (reserved field, currently unused)
system_prompt	System prompt
input_limit	Maximum number of characters for user input, 0 means no limit

⚠️ The config.yaml parser is a minimalist implementation, only supporting

key: value format (one key per line)

---

🚀 Quick Start

1️⃣ Compile and Run

go build -o agent-code-assistant main.go

./agent-code-assistant

2️⃣ Input Example

> List files in the current directory

> Read the contents of main.go

> Create a file hello.txt with the content "Hello World"

LLM If it returns the following format:

tool: read_file({"path":"main.go"})

The program will automatically execute the tool and continue the conversation.

---

🧰 Built-in Tools

Currently, there are 3 built-in tools (can be freely expanded):

Tool Name	Function
read_file	Read file content
list_files	List directory files
edit_file	Create/modify file

Tool Call Format (Convention)

tool: tool_name({"arg":"value"})

Example:

tool: edit_file({"path":"a.txt","old":"","new":"hello"})

---

🔐 Input Length Limit

Controlled via config.yaml:

input_limit: 200

Behavior:

• 0 → No limit (default)

• >0 → Input exceeding the character limit will be rejected.

Character counting uses rune, which is safe for Chinese characters.

---

📌 Who is it suitable for?

• For those who want to understand the essence of AI programming assistants

• For those who want to implement the core logic of Copilot/Claude Code themselves

• For engineers who want to create private/localized/auditable AI tools

• For those dissatisfied with the "black box" nature of Agent frameworks

---

🧪 Non-goals (Intentionally Avoided)

• ❌ Not pursuing production-grade robustness

• ❌ Not implementing complex prompt templates

• ❌ Not having a built-in permission system/sandbox

• ❌ Not introducing Agent/Workflow frameworks

---

🛠 Scalability Directions (Suggestions)

• Add max_steps to prevent infinite loops

• Add exec_cmd (restricted shell)

• Support multi-tool JSON arrays

• Context persistence (session recovery)

• Transform into an MCP Server (Go)

---

📜 License

MIT. You are free to modify, tailor, rewrite, and use it commercially.