Aider

Open-source CLI AI coding assistant. Architect mode uses two models — one designs the solution, one applies file edits — beating single-model code mode for complex multi-file changes. Repo map gives the LLM full codebase context via tree-sitter.

Updated Invalid Date·

aider coding-assistant cli repo-map architect-mode editor-mode open-source

The open-source CLI coding assistant most comparable to Claude Code for whole-repo editing. Aider works inside any terminal, supports any Git repo, creates atomic commits per change, and lets you pair LLMs: a powerful "architect" model reasons about what to change while a cheaper/faster "editor" model applies the file edits.

[Source: aider.chat, 2026-05-03]

Key Facts

Website: https://aider.chat
Install: pip install aider-chat
License: Apache 2.0
Git integration: creates atomic commits with co-authored-by attribution
Repo map: tree-sitter based; indexes function signatures, class definitions, and file structure
Models supported: Claude, GPT-4o, Gemini, DeepSeek, local models via Ollama
Benchmarks: Aider Polyglot leaderboard (225 exercises across C++, Go, Java, JS, Python, Rust)

Modes

Code mode (default)

One model proposes the solution and applies file edits. Works well for straightforward single-file changes.

aider --model claude-opus-4-7 app.py
> Add input validation to the create_user function

Architect mode

Two models collaborate: an architect designs the solution; an editor applies the edits to files. Significantly better for complex, multi-file changes where reasoning and editing are separate skills.

aider --architect --model claude-opus-4-7 --editor-model claude-haiku-4-5
> Refactor the auth module to use JWT instead of session tokens

The architect model sees the full context (repo map + relevant files) and proposes the changes in natural language. The editor model converts the proposal into precise file edits using editor-diff or editor-whole format.

When to use architect mode:

Multi-file refactors
Architecture changes spanning multiple modules
When the default model produces incomplete or incorrect edits
When you want to use a cheaper editor model to reduce cost

Ask mode

Chat without modifying files — for exploring options, understanding code, or getting a plan before committing to changes.

aider --ask
> What's the best approach to add rate limiting to this API?

Watch mode

Aider monitors your files for AI comments (e.g., # AI: add type hints to this function) and triggers automatically. Useful for annotation-driven development.

Repo Map

Aider builds a map of your entire repository using tree-sitter:

Extracts function signatures, class definitions, method names
Compresses to fit in the LLM's context window
Prioritises files most relevant to the current task
Updated dynamically as files change

This gives the LLM awareness of the whole codebase without requiring it to read every file. The repo map is the key reason Aider can propose cross-file changes coherently.

Git Workflow

Every change Aider makes is committed as an atomic commit:

Add input validation to create_user

Co-Authored-By: aider (claude-opus-4-7) <aider@example.com>

Benefits:

Easy to git diff HEAD~1 to review what Aider did
Easy to git revert if the change is wrong
Clean history showing which commits were AI-assisted

Configuration

# ~/.aider.conf.yml
model: claude-opus-4-7
architect: true
editor-model: claude-haiku-4-5-20251001
auto-commits: true
cache-prompts: true   # prompt caching for Anthropic models

Or via environment:

export ANTHROPIC_API_KEY=sk-ant-...
aider --model claude-sonnet-4-6 --architect

Integrations

/web: fetch and include a URL's content in context (/web https://docs.example.com)
/voice: voice input via whisper
/paste: paste clipboard content into context
/run: run shell commands and add output to context
MCP servers: Aider does not natively consume MCP tools. Third-party community projects (e.g., aider-mcp-server) expose Aider as an MCP server for other clients to invoke.

vs Claude Code

Dimension	Aider	Claude Code
Interface	CLI (any terminal)	CLI + IDE extensions
Autonomy	Lower — proposes, you accept	Higher — executes shell commands
Repo map	Built-in (tree-sitter)	CLAUDE.md + dynamic context
Git commits	Automatic (atomic)	Manual or on request
Model flexibility	Any model (multi-provider)	Claude only
Cost control	Architect+editor split	Single model
Best for	Terminal-native devs, model experimentation	Autonomous multi-step tasks

Connections

ai-tools/claude-code — Claude Code is the closest equivalent from Anthropic; higher autonomy, Claude-only
ai-tools/cursor-copilot — Cursor is the IDE-integrated alternative; less CLI-native
apis/anthropic-api — Aider uses the Anthropic SDK for Claude models; prompt caching supported
prompting/context-engineering — repo map is a form of context engineering at the tool layer

Open Questions

Does Aider's architect mode outperform Claude Code's agent mode on SWE-bench tasks?
How does watch mode interact with CI/CD pipelines for annotation-driven development?
What is the repo map size limit before tree-sitter compression degrades quality?