Aider AI Coding

Terminal-based AI coding with Aider on Clore.ai — Git-aware, multi-file, local models via Ollama

Aider is a terminal-based AI coding assistant with 39K+ GitHub stars. It edits files directly in your repo, creates Git commits automatically, and supports both cloud APIs (OpenAI, Anthropic) and fully local models via Ollama. On a Clore.ai GPU, you can run large coding models like DeepSeek-R1 32B or Qwen2.5-Coder-32B entirely on your own hardware — private, fast, and cost-effective.

All examples run on GPU servers rented through the CLORE.AI Marketplace.

Key Features

• Terminal-native — works over SSH, perfect for headless Clore.ai servers

• Git-aware — auto-commits every change with descriptive messages, easy to review and revert

• Multi-file editing — add multiple files to context and edit them simultaneously

• Local model support — connect to Ollama for fully private, zero-API-cost coding

• Architect mode — use a strong reasoning model to plan, then a fast model to implement

• Repository map — automatically indexes your codebase for context-aware edits

• Linting and testing — run linters/tests after each edit, auto-fix failures

• Voice input — dictate coding instructions via microphone

Requirements

Component	Minimum	Recommended
GPU	RTX 3060 12 GB	RTX 4090 24 GB
VRAM	12 GB	24 GB
RAM	16 GB	32 GB
Disk	30 GB	60 GB
Python	3.9	3.11

Clore.ai pricing: RTX 4090 ≈ $0.5–2/day · RTX 3090 ≈ $0.3–1/day · RTX 3060 ≈ $0.15–0.3/day

For cloud-only models (no local inference), a GPU is not required — but Clore.ai GPUs let you run Ollama models locally for full privacy.

Quick Start

1. Install Aider

pip install aider-chat

2. Set Up Ollama for Local Models

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull a coding model
ollama pull deepseek-r1:32b

# Or a smaller model for lower VRAM
ollama pull qwen2.5-coder:7b

3. Start Aider with a Local Model

cd /workspace/your-project

# Use DeepSeek-R1 32B via Ollama (needs ~20 GB VRAM)
aider --model ollama/deepseek-r1:32b

# Or use a smaller model on RTX 3060
aider --model ollama/qwen2.5-coder:7b

4. Start Coding

Inside the Aider REPL:

> /add src/main.py src/utils.py
> Add error handling to the parse_config function and write unit tests for it

Aider will:

1. Read the files and understand the codebase

2. Propose changes as a diff

3. Apply the changes to disk

4. Create a Git commit with a descriptive message

Usage Examples

Architect Mode (Two-Model Setup)

Use a strong model for reasoning and a fast model for code generation:

# Cloud architect + local editor
aider --architect --model ollama/deepseek-r1:32b --editor-model ollama/qwen2.5-coder:7b

The architect model plans the changes, and the editor model writes the actual code — combining high-quality reasoning with fast implementation.

Add Files and Edit

# Add specific files to the chat context
> /add src/api/routes.py src/models/user.py

# Ask for changes
> Refactor the user registration endpoint to use async/await and add input validation with Pydantic

# Add a whole directory
> /add src/tests/

# Run tests after the edit
> /test pytest src/tests/ -v

Use with Cloud APIs

# OpenAI
export OPENAI_API_KEY=sk-...
aider --model gpt-4o

# Anthropic
export ANTHROPIC_API_KEY=sk-ant-...
aider --model claude-sonnet-4-20250514

Git Integration

# Every change creates a commit automatically
git log --oneline -5
# a1b2c3d aider: Add error handling to parse_config
# d4e5f6g aider: Write unit tests for parse_config
# h7i8j9k aider: Refactor user registration endpoint

# Undo the last aider change
aider --undo

Lint and Auto-Fix

# Configure a linter
aider --lint-cmd "ruff check --fix" --auto-lint

# Aider runs the linter after each edit and auto-fixes issues

Non-Interactive (Scripted) Mode

# Run a single instruction and exit
aider --model ollama/deepseek-r1:32b \
  --message "Add type hints to all functions in src/utils.py" \
  --yes  # auto-accept changes

Model Recommendations

Model	VRAM	Speed	Quality	Best For
deepseek-r1:32b	~20 GB	Medium	High	Complex refactoring
qwen2.5-coder:32b	~20 GB	Medium	High	Code generation
qwen2.5-coder:7b	~5 GB	Fast	Good	Quick edits, RTX 3060
codellama:34b	~20 GB	Medium	Good	Legacy code, C/C++
deepseek-coder-v2:16b	~10 GB	Fast	Good	Balanced performance

Tips

• Use /add selectively — only add files Aider needs to see. Too many files waste context tokens

• Architect mode is powerful for complex changes — the reasoning model catches edge cases the editor model might miss

• /undo reverts the last change cleanly via Git — experiment freely

• /diff shows the proposed changes before applying — use for review

• Set --auto-commits (default) for full Git history of every AI change

• Use .aiderignore to exclude files from the repo map (node_modules, .venv, etc.)

• For large repos, Aider's repo map helps the model understand code structure — let it run on first load

• Run tests after edits — /test pytest catches regressions immediately

Troubleshooting

Problem	Solution
Ollama model too slow	Use a smaller quantization (q4_0) or smaller model
CUDA out of memory with Ollama	Pull a smaller model variant or use OLLAMA_NUM_GPU=0 for CPU
Git commit errors	Ensure git config user.email and user.name are set
Aider ignores my files	Use /add filename.py explicitly — Aider only edits added files
Model produces poor edits	Try a stronger model, or use architect mode
Connection refused (Ollama)	Ensure Ollama is running: ollama serve or systemctl start ollama
Context window exceeded	Remove files with /drop, keep only relevant ones

Resources

• Aider GitHub

• Aider Documentation

• Ollama Model Library

• CLORE.AI Marketplace