AI Coding Agent

2024 – Present · Active

Overview

AI Coding Agent is a local-first, multi-agent coding assistant that plans work, writes and refines code, debugs failures, and uses tools against your filesystem and runtime. It runs on a local LLM via Ollama (e.g. DeepSeek Coder), so prompts and generated code stay on your machine—ideal when you want automation without sending proprietary repos to a hosted API.

The implementation follows a specialized multi-agent pattern: separate agents focus on planning, implementation, and debugging, coordinated through a core orchestrator and a tools layer (file ops, execution helpers, etc.). Long-horizon context is supported with structured memory (Supabase) and semantic recall via Chroma vector memory, so later steps can still “remember” earlier decisions and codebase facts.

Repository: Milanirosefrancis/ai-coding-agent — clone it to run the CLI, configure Ollama, and wire optional memory backends from requirements.txt.

What it can do

  • End-to-end project work — describe a feature or greenfield task; the planner breaks it down, the coder generates patches, and the debugger tightens loops when something breaks.
  • GitHub-style inputs — point the workflow at repositories or unpacked code so analysis and edits stay grounded in real project layout.
  • Links and external context — fetch and reason about content from URLs when you need API docs, specs, or reference material alongside the repo.
  • Autonomous tool use — the agent doesn’t only “chat”; it can execute allowed tools (per project safety rules) to read files, run commands, and iterate.

Architecture (mental model)

  1. CLI — you describe intent, approve or steer plans, and inspect logs.
  2. Core agent — routes work to specialists and merges their outputs.
  3. Planner / Coder / Debugger — narrow prompts and responsibilities reduce confusion and help the local model stay on task.
  4. Memory — Supabase for durable session or project state; Chroma for embeddings-backed retrieval over notes, prior outputs, or indexed fragments.
  5. Ollama — inference for all LLM turns; swap models by changing your Ollama tag.

Why it’s interesting

Shipping a multi-agent system on a local stack forces clear boundaries between planning, coding, and verification—and makes memory + retrieval first-class instead of an afterthought. This repo is a solid base for experiments in agentic workflows, RAG over dev artifacts, and tool-grounded code generation without cloud lock-in.

Architecture

High-level view of how the React client, Express API, data stores, and phased AI services fit together.

Tech stack

Python
Ollama · DeepSeek Coder
Multi-agent architecture
Supabase (memory)
Chroma · vector memory
Flask API
CLI
Automatic tool execution
GitHub & repo analysis
Web / link analysis