---

CodingAgent

CodingAgent is an open-source, CLI-based AI coding agent that works like the best closed-source tools — Claude Code, Codex, and Antigravity — but without the vendor lock-in. It supports any LLM (cloud or local), persists memory across sessions, integrates natively with Git and MCP servers, and stays lightweight enough for real engineering workflows.

---

🚀 Features

• Model Agnostic LLM Support: Works with any LLM out of the box — Claude, GPT, Gemini, Ollama local models, Kimi — switched via a single config line with no code changes required.
• Persistent Multi-Tier Memory: Three-tier memory system (short-term buffer, long-term local vector store, per-project AGENT.md scratchpad) that survives session resets and context limits — the agent remembers decisions, patterns, and past bugs.
• Reliable Agent Loop with File & Shell Tools: A clean goal → plan → execute → observe loop with native file read/write, shell command execution, and automatic error recovery — no bloated orchestration framework, just a tight and predictable core loop.
• MCP Integration: First-class Model Context Protocol support so users can plug in GitHub, databases, documentation, or any custom tool server via a simple config entry — no core code changes needed.
• Optimised Context & File Caching: Files already read in a session are cached and not re-fetched unless modified — drastically cuts token usage and API costs across long sessions.
• User-Editable Prompt Profile: A simple profile.md where users describe their tech stack, coding style, and preferences — the agent reads this on every session startup and codes exactly the way you like.
• Git Integration: Native Git awareness — reads commit history and diffs so the agent understands what changed recently and makes smarter, context-aware suggestions over time.
• Commit-Level Code Review: On-demand security and quality review for any commit or staged diff — lists vulnerabilities, code smells, and risky changes as a clean CLI report when explicitly asked.

---

💻 Tech Stack

CLI & Runtime

• TypeScript
• Commander.js (argument parsing)
• Inquirer (interactive terminal prompts)

Backend / Core Agent

• Node.js
• Execa (safe async shell execution)
• Ripgrep + Glob (fast codebase file navigation)
• Simple Git (native Git integration — diffs, history, commits)

Database & Memory

• SQLite (local persistent storage, zero external services)
• sqlite-vec (vector search extension for semantic memory)
• In-memory JSON ring buffer (short-term session context)

AI/ML

• Vercel AI SDK (provider-agnostic LLM routing)
• Google Gemini / OpenAI GPT / Anthropic Claude / Ollama / Kimi
• Model Context Protocol SDK (MCP tool + server connections)
• Custom token budget manager + auto-compaction engine
• RAG over local codebase / user prompt profile engineering

Tooling & Build

• tsup (TypeScript bundler)
• tsx (zero-config TypeScript dev runner)
• Vitest (unit and integration testing)
• ESLint + Prettier (code quality and formatting)
• SWE-bench (agent performance benchmarking)

---

✅ Project Checklist

• The CLI Agent (core):

  • Agent loop (goal → plan → execute → observe) is implemented and documented.
  • Works end-to-end with at least one local model (Ollama) and one cloud model (Claude/GPT).
  • Installable via npx codingagent with zero manual setup.

• Multi-LLM Support:

  • Provider-agnostic LLM routing implemented and tested across Claude, OpenAI, Gemini, and Ollama.
  • Model switching works via single config change with no code modification.
  • Streaming responses work correctly across all supported providers.

• Memory System:

  • Short-term in-session buffer implemented with token cap.
  • Long-term SQLite vector store persists across sessions.
  • Per-project AGENT.md scratchpad is read and written by the agent automatically.
  • File caching prevents re-reading unchanged files within a session.

• Context & Cost Optimisation:

  • Token budget manager tracks and allocates context per turn.
  • Auto-compaction triggers when context hits 80% of model limit.
  • User-editable profile.md for tech stack and coding style preferences is loaded on every session.

• MCP Integration:

  • MCP client connects to at least GitHub and one custom tool server.
  • New MCP servers can be added via config with no code changes.

• Git Integration:

  • Agent reads commit history and diffs natively.
  • On-demand commit-level code review lists vulnerabilities and risky changes in CLI.

• The AI/ML components:

  • LLM model selection and configuration are documented.
  • System prompt and user profile.md are version-controlled.
  • API keys are stored securely in local config, never hardcoded.
  • Rate limit handling and retry logic implemented for all providers.

• Quality & Open Source:

  • Test coverage for core agent loop, memory, and tool execution.
  • Benchmarked against SWE-bench or equivalent to prove quality.
  • Contributing guide and onboarding docs published in repo.

---

🔗 Repository Links

1. Main Repository
2. GSoC 2026 Idea
3. GSoC 2026 Index

---

🏗️ Architecture Diagram

---

🔄 User Flow

---

�🍀 Getting Started

To be added.

---

📱 App Screenshots

To be added.

---

🙌 Contributing

⭐ Don't forget to star this repository if you find it useful! ⭐

Thank you for considering contributing to this project! Contributions are highly appreciated and welcomed. To ensure smooth collaboration, please refer to our Contribution Guidelines.

---

✨ Maintainers

• Himanshu Poptani

---

📍 License

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.

---

💪 Thanks To All Contributors

Thanks a lot for spending your time helping CodingAgent grow. Keep rocking 🥂

© 2026 AOSSIE