Show HN: Elelem, a tool-calling CLI for Ollama and DeepSeek in C

4 months ago 3

Chat Client with Tool-Calling Support

An interactive C-based command-line chat client that enables AI models to execute real-world actions through a comprehensive tool-calling system.

🚀 Quick Start

Prerequisites

Ubuntu/Debian:

sudo apt-get update sudo apt-get install build-essential libglib2.0-dev libjson-glib-dev libsoup2.4-dev libreadline-dev

Fedora/RHEL:

sudo dnf install gcc glib2-devel json-glib-devel libsoup-devel readline-devel

macOS:

brew install glib json-glib libsoup readline pkg-config

Building

git clone <your-repo-url> cd gobject make clean && make

Setup API Keys

For DeepSeek API:

export DEEPSEEK_API_KEY="sk-your-key-here"

For Ollama (local):

# Install and start Ollama curl -fsSL https://ollama.ai/install.sh | sh ollama serve # Pull a model (in another terminal) ollama pull qwen3

📖 Usage

Basic Commands

# Start with DeepSeek (requires API key) ./elelem # Use Ollama with specific model ./elelem -p ollama -m qwen3 # Use custom Ollama server ./elelem -p ollama -u http://remote-server:11434 -m mistral

Interactive Commands

Command Description Example

/tools	List available tools	/tools
/history	Show conversation history	/history
/save	Save current conversation	/save
/load <file>	Load previous conversation	/load conversations/chat_2024-01-01_10-30-00.txt
/model <name>	Switch AI model	/model llama3.1
/provider <name>	Switch between providers	/provider ollama
/clear	Clear conversation history	/clear
/exit	Exit application	/exit

🛠️ Available Tools

The AI can execute these tools to help you:

🔍 Search Tools

grep - Search for patterns in files recursively
- Required: pattern
- Optional: path, file_pattern, case_sensitive
analyze_code - Analyze code structure and metrics
- Optional: path (default: current), language (default: c)

📁 File Tools

read_file - Read and display file contents
- Required: filename
write_file - Create or modify files
- Required: filename, content
list_directory - Browse directory contents
- Optional: path (default: current directory)

💻 System Tools

shell - Execute shell commands (sandboxed for safety)
- Required: command

💡 Example Interactions

Code Analysis

> Analyze this codebase and tell me about its structure [AI will use the analyze_code tool to examine your project, count lines of code, identify functions, and provide architectural insights]

File Search

> Find all TODO comments in C files [AI will use grep tool with pattern "TODO" and file_pattern "*.c" to search recursively through your codebase]

File Operations

> Create a new header file called "utils.h" with basic includes [AI will use write_file tool to create the header file with appropriate content]

System Operations

> What's in the current directory and what's the git status? [AI will use list_directory and shell tools to show directory contents and run "git status"]

🏗️ Architecture

Core Components

main.c - CLI interface and main event loop
llm_interface.[ch] - Abstract client interface
deepseek_client.[ch] - DeepSeek API implementation
ollama_client.[ch] - Ollama local API implementation
tool_manager.[ch] - Tool orchestration system
tool_definition.[ch] - Tool schema definitions
builtin_tools.c - Built-in tool implementations
my_http_client.[ch] - HTTP client utilities

Tool-Calling Flow

User Input → User asks AI to perform a task
AI Planning → Model decides which tools to use
Tool Calls → AI generates JSON function calls
Validation → System validates tool calls and parameters
Execution → Tools run in sandboxed environment
Results → Tool output fed back to AI
Response → AI provides final answer with context

🔧 Configuration

System Prompt Customization

The system prompt is loaded from prompts/system_prompt.txt. You can customize it to:

Add new tool descriptions
Modify AI behavior
Change response format preferences

Conversation Storage

Conversations are auto-saved to conversations/ directory
Files are named with timestamps: chat_YYYY-MM-DD_HH-MM-SS.txt
Use /load command to resume previous conversations

🛡️ Security Features

Command Filtering - Dangerous shell commands are blocked
Path Validation - File operations validate and sanitize paths
Output Limits - Large outputs are truncated to prevent memory issues
Sandboxed Execution - Tools run with limited permissions
Input Validation - All tool parameters are validated before execution

🔨 Development

Adding New Tools

Implement the handler in builtin_tools.c:

static ToolResult * my_custom_tool_handler (JsonObject *arguments, gpointer user_data) { // Your implementation here ToolResult *result = g_new0 (ToolResult, 1); result->success = TRUE; result->content = g_strdup ("Tool output"); return result; }

Register the tool in tool_manager_register_builtin_tools():

tool = tool_definition_new ("my_custom_tool", "Description of what it does"); tool_definition_add_string_param (tool, "param_name", "Parameter description", TRUE); tool_manager_register_tool (manager, tool, my_custom_tool_handler, NULL, NULL);

Update system prompt in prompts/system_prompt.txt to describe the new tool.

Building with Debug Info

make clean CFLAGS="-g -O0 -DDEBUG" make

Running Tests

# Test with different providers ./elelem -p deepseek ./elelem -p ollama -m llama3.1 # Test tool functionality echo "List the files in this directory" | ./elelem -p ollama -m llama3.1

🐛 Troubleshooting

Common Issues

Build Errors:

Ensure all dependencies are installed
Check pkg-config can find libraries: pkg-config --cflags glib-2.0

DeepSeek API Issues:

Verify API key is set: echo $DEEPSEEK_API_KEY
Check network connectivity and API quotas

Ollama Issues:

Ensure Ollama server is running: curl http://localhost:11434/api/version
Verify model is available: ollama list

Tool Execution Issues:

Check file permissions for file operations
Verify shell commands aren't blocked by security filters

Debug Mode

Set environment variable for verbose output:

G_MESSAGES_DEBUG=all ./elelem

📊 Features

✅ Implemented

Multi-provider support (DeepSeek, Ollama)
Real-time streaming responses
Tool-calling with 6 built-in tools
Conversation history with save/load
Command-line interface with readline support
Security sandboxing and validation
File-based system prompt configuration

🚧 Planned

Tool result caching for performance
Async tool execution for parallel operations
Plugin system for third-party tools
Tool dependency management and chaining

📄 License

GNU Affero General Public License v3

Need help? Open an issue or check the conversation history with /history to see example interactions.

Read Entire Article