AI Coding Tool
Aider
The open-source terminal coding assistant where every edit is a git commit
Last updated: April 15, 2026
Quick verdict
Aider is best for developers who want open-source AI coding with a git-native workflow and full model flexibility — every edit is a commit, so git revert is your undo button. Supports 100+ LLMs via LiteLLM, including local models through Ollama for offline or privacy-sensitive work. Architect mode lets Claude Opus 4 plan while GPT-4o-mini implements, cutting cost 10x. No GUI, no subscription, no seat fee — you pay only your API provider.
Model
Claude Sonnet 4 / GPT-4o / Gemini / Ollama local models (100+ LLMs)
From
$0/mo
Free tier
Yes
Setup
5 min
Best for
- + Developers who want open-source AI tooling with no vendor lock-in
- + Git-first workflows where every change must be a reviewable commit
- + Privacy-sensitive code via local models through Ollama
- + BYOK users wanting a choice of 100+ LLM providers
- + Cost-conscious developers using architect mode to save 10x on token costs
- + Terminal-only developers who do not want a GUI IDE
Not for
- - Developers who want a visual IDE with inline completions
- - Teams wanting a managed SaaS with SSO and audit logs
- - Beginners unfamiliar with terminal and git workflows
- - Projects needing built-in codebase-wide semantic search
Key features
Every edit is a git commit
UniqueAider commits each AI edit to git automatically with a descriptive message. Undo any change with git revert HEAD or git reset. The entire history is auditable — you see exactly what the AI changed and when.
Architect mode: plan with one model, code with another
UniqueUse an expensive model (Claude Opus 4, o1) to plan the change, then a cheaper model (GPT-4o-mini, Claude Haiku) to implement. Cuts cost 5-10x on complex tasks while keeping reasoning quality high.
100+ LLM support via LiteLLM
UniqueClaude, GPT, Gemini, Cohere, Mistral, DeepSeek, Groq, Fireworks, and dozens more via LiteLLM. Any OpenAI-compatible endpoint works. Configure with environment variables — no vendor lock-in.
Local and offline via Ollama
UniqueRun Aider fully offline with Ollama-hosted models — DeepSeek Coder, Qwen Coder, Code Llama, Devstral. No code or prompts leave your machine. Slower than hosted models but free and private.
Voice input mode
Speak prompts instead of typing. Aider transcribes via OpenAI Whisper and sends to your chosen coding model. Useful for long prompts or hands-free work.
Repo map with tree-sitter
Aider builds a compact symbol map of your repo using tree-sitter so the model sees relevant function signatures and types across files without loading every file into context. Keeps token costs down on large codebases.
Limitations
Terminal-only — no GUI
Aider runs in the terminal. There is no IDE integration, no visual diff panel, no file tree. Review diffs with git diff or your editor's source control panel.
Smaller community than Cursor or Copilot
Aider has strong open-source community on GitHub but fewer tutorials, templates, and third-party integrations than Cursor, Copilot, or Claude Code.
Architect mode adds latency
The two-model pipeline (architect plans, editor implements) doubles round trips. A task that takes 15 seconds with a single model may take 30-40 seconds in architect mode. Worth it for hard tasks, overkill for simple edits.
No built-in codebase semantic search
Aider's repo map gives structural context (symbols, signatures) but there is no embedding-based semantic search over the whole codebase. You must /add relevant files manually for tasks that span many files.
Benchmarks
Tips
Use architect mode for hard tasks, single-model for easy ones
Architect mode (--architect with Claude Opus 4 + GPT-4o-mini editor) cuts cost 5-10x on complex multi-file work. For single-file edits or simple refactors, a single Claude Sonnet 4 call is faster and good enough.
Use /add and /drop to manage context
Aider only sends files you /add to the model. Keep the working set small — 3 to 8 files typically. /drop files when you are done with them. This keeps token costs low and focuses the model on relevant code.
Try a local Ollama model for sensitive code
For confidential repos, run Aider against a local Ollama server with Qwen Coder 32B or DeepSeek Coder. Slower than Claude Sonnet 4 but nothing leaves your machine. Set OLLAMA_API_BASE and pick the model with --model.
Let Aider's auto-commits be your safety net
Every edit is already a git commit. Experiment freely — if an edit is wrong, git reset HEAD~1 puts you back. Treat Aider sessions as branchable experiments rather than risky direct edits.
Frequently asked questions
How much does Aider cost?
Aider itself is free and open-source (Apache 2.0 license). You pay your chosen API provider directly — Anthropic, OpenAI, Google, or others — or run local models via Ollama for zero API cost. Typical daily use with Claude Sonnet 4 runs $2-6/day.
Why does every Aider edit become a git commit?
Aider auto-commits each AI edit with a descriptive message so you have a full audit trail and an easy undo path. Wrong edit? git reset HEAD~1 or git revert HEAD. The git-native workflow is Aider's core differentiator versus tools that apply edits directly to the working tree.
Which models does Aider support?
Aider supports 100+ LLMs via LiteLLM: Claude Sonnet 4, Claude Opus 4, GPT-4o, o1, o1-mini, Gemini 2.0 Pro, DeepSeek V3, Mistral Large, Cohere Command R+, Groq-hosted models, and any OpenAI-compatible endpoint. Local models work via Ollama — Qwen Coder, DeepSeek Coder, Code Llama, Devstral.
How does Aider's architect mode work?
Architect mode uses two models: an expensive reasoning model (Claude Opus 4 or o1) to plan the change, and a cheaper model (GPT-4o-mini, Claude Haiku) to implement the plan in code. Typically cuts cost 5-10x on complex multi-file work while preserving reasoning quality. Enable with --architect.
How does Aider compare to Claude Code?
Both are terminal-native agentic coding tools. Claude Code is closed-source and Anthropic-only with hooks, MCP, and subagents. Aider is open-source and supports 100+ models including local ones via Ollama, with a git-commit-per-edit workflow. Aider wins on flexibility and cost; Claude Code wins on agentic depth, hooks, and ecosystem.
Can Aider run fully offline with local models?
Yes. Run an Ollama server locally with a coding model like Qwen Coder 32B or DeepSeek Coder, then point Aider at it with OLLAMA_API_BASE and --model ollama/qwen2.5-coder:32b. No code or prompts leave your machine. Slower than hosted Claude or GPT-4o but free and fully private.