This is a project I built with codecrafters (https://app.codecrafters.io/) with their Build-your-own-claude-code challenge , but there are many extensions to it now in this project . Only tested with GROQ provider , so please open an issue if you encounter any problems with other providers.

Version 0.1.0

Claude Code C - AI-Agnostic Coding Assistant

Claude Code C is an advanced, terminal-based AI coding assistant written in C (using the C23 standard). Inspired by Anthropic's Claude Code and Agy, it features a complete multi-turn agent loop, interactive REPL, an automated tool execution system, and multi-provider compatibility.

This project is AI provider-agnostic, meaning it can interface with OpenRouter, OpenAI, Anthropic, Google Gemini, Groq, or local Ollama servers. It translates payloads transparently between standard OpenAI Chat Completions formats and native Anthropic Messages API formats under the hood.

---

Key Features

Feature	Description	Supported Backends
Interactive REPL	Multi-turn chat interface maintaining full session conversation history.	All
Interactive Model Dropdown	Choose providers and models interactively on startup, or switch dynamically using /settings and /model.	All
Tool Calling Suite	7 pre-configured tools enabling the LLM to inspect, read, search, and edit the filesystem.	All
Provider Agnostic	Standard API client for multiple AI services and local LLMs.	OpenAI, Anthropic, Gemini, Groq, Ollama, OpenRouter
Anthropic Adapter	In-memory translation layer converting OpenAI formats into Anthropic Messages schemas.	Anthropic
Custom System Prompts	Prepend custom rules combined with default system prompt instructions.	All
Resilient Agent Loop	Automatic retries and feedback loops providing tool error output back to the LLM for correction.	All

---

Tool Calling Suite

Claude Code C implements the following tools that the AI can call dynamically to explore codebases, inspect environment metadata, or perform writes and commands:

Tool Name	Parameters	Description
Read	file_path	Reads and returns the contents of a file on the local filesystem.
Write	file_path, content	Writes or overwrites contents of a file. Automatically creates directories.
Bash	command	Runs a shell script/command inside an interactive Bash environment and returns output.
ListDirectory	path (optional)	Lists files, links, and directories inside the given path.
FindFiles	directory, pattern	Recursively walks directories to locate files matching the pattern.
FileSearch	directory, query	Recursive search for text occurrences inside files (similar to grep).
SysInfo	None	Returns system details including OS, machine type, user, CWD, and shell.

---

Compilation and Setup

Prerequisites

• A C compiler supporting C23 (GCC 13+ or Clang 16+)
• CMake (version 3.13+)
• Libcurl and cJSON library. If using vcpkg, dependencies are handled automatically via vcpkg.json.

Building

Compile using CMake:

cmake -B build -S . -DCMAKE_TOOLCHAIN_FILE=${VCPKG_ROOT}/scripts/buildsystems/vcpkg.cmake
cmake --build ./build

Alternatively, use the convenience wrapper script:

./your_program.sh --help

---

Command Line Usage

Option (Short)	Option (Long)	Argument	Description
-p	--prompt	string	Runs a single user prompt non-interactively and exits.
-i	--interactive	None	Starts an interactive REPL session (default).
-c	--choose-model	None	Launches interactive dropdown menus to choose provider, model, and input API key.
-r	--provider	provider	Selects LLM backend: openrouter, openai, anthropic, gemini, groq, ollama.
-m	--model	string	LLM model identifier.
-k	--api-key	string	API token/secret key.
-u	--base-url	string	Custom API endpoint URL.
-s	--system-prompt	file	Path to a file containing custom system prompt rules.
None	--user-system-prompt	string	String containing custom system prompt rules directly.
-h	--help	None	Prints the help menu and lists options.

---

Environment Variables

You can configure the application without command-line arguments using these environment variables:

Variable	Description	Default (per provider)
AI_PROVIDER	AI service provider (openai, anthropic, gemini, groq, ollama, openrouter).	openrouter
AI_MODEL	The LLM model to request.	Provider specific (e.g. gpt-4o, claude-3-5-sonnet-20241022)
AI_API_KEY	API secret key (falls back to provider-specific environment variables).	None
AI_BASE_URL	Override endpoint for custom hosts or local deployments.	Provider specific (e.g. http://localhost:11434/v1 for Ollama)
AI_SYSTEM_PROMPT_FILE	Path to file containing custom instructions.	None

---

Architecture and Implementation Details

                    +------------------------------------+
                    |             main.c                 |
                    |   (CLI Parser & Env configuration)  |
                    +-----------------+------------------+
                                      |
                                      v
                    +------------------------------------+
                    |             app.c                  |
                    |   (Interactive REPL & Agent Loop)   |
                    +--------+------------------+--------+
                             |                  |
                             v                  v
              +--------------+----+        +----+--------------+
              |    provider.c     |        |  tool_registry.c  |
              |                   |        |                   |
              | (cURL HTTP Client |        |  (Read, Write,    |
              |  Anthropic Schema |        |   Bash, Grep,     |
              |  Translation      |        |   Find, SysInfo)  |
              +-------------------+        +-------------------+

1. In-Memory Request & Response Translation

Anthropic's Messages API specifies system prompts, tool schemas, and assistant replies differently than the standard OpenAI Chat Completions API. Instead of duplicating the agent loop, provider.c houses a dual translation layer:

• Request Translator: Converts messages by moving system messages to top-level fields, renaming parameter fields to input_schema, and combining contiguous tool call messages.
• Response Translator: Reconstructs standard OpenAI tool call formats from Anthropic's output structure.

2. Multi-turn History and Interactive Agent Loop

The REPL is powered by a JSON messages array that grows during the session. When the LLM issues tool calls, the agent loop pauses, invokes the tool registry, displays status updates to the terminal, and posts the responses back to the LLM. The agent runs up to 20 sub-iterations dynamically before yielding to the user.