Aider Deep Dive

★ ★ Aider (github.com/Aider-AI/aider, Apache 2.0) is the dominant FOSS terminal-based AI pair programmer in 2026. Started by Paul Gauthier; widely used; works with any OpenAI-compatible model — Ollama, OpenRouter, Anthropic, OpenAI, DeepSeek, Mistral. Diff-based edits with human-in-loop review. Repo-map for codebase awareness. Architect / Editor mode for splitting reasoning and execution.

If you're shopping for ★ ★ "the local-model agentic coding pick," this is the answer. See agentic coding overview, coding models, hardware tiers.

What Aider is

• ★ Terminal-based. Run aider in your project directory; chat with the model about code.
• ★ Diff-based edits. The model produces search/replace blocks; Aider applies them. You review every diff before commit.
• ★ Auto-commits to git. Each accepted change is a commit. Easy revert.
• ★ Repo map. Aider sends a tree-sitter-derived skeleton of your codebase as context, so the model knows what exists without reading every file.
• ★ Voice mode. Whisper-based dictation for hands-free coding.
• ★ Architect mode. Reasoning model (e.g., R1) plans; faster code model executes.
• ★ Multi-file edits. One prompt can edit several files coherently.
• ★ Test integration. --test-cmd "pytest" runs your tests after each edit; the model sees failures and fixes.
• ★ Linter integration. Same idea with lint output.
• License: Apache 2.0.

Quick start

# Install
pip install aider-install
aider-install
 
# Local with Ollama
export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama/qwen2.5-coder:32b
 
# OpenRouter (mix of providers)
export OPENROUTER_API_KEY=...
aider --model openrouter/anthropic/claude-sonnet-4.7
 
# Hybrid: architect + editor
aider \
  --architect --model ollama/deepseek-r1-distill-llama:70b \
  --editor-model ollama/qwen2.5-coder:32b

Modes

Code (default)

• Direct edits via search/replace blocks.
• Pair-programming-style turn-by-turn.

Architect

• The architect model reasons / writes a plan in plain prose.
• The editor model translates the plan into actual diffs.
• ★ Big win for hard refactors when you have a smart-but-slow reasoning model and a fast-and-precise code model.

Ask

• Q&A only; no edits proposed.

Help

• Aider documentation and command reference.

Model picks for Aider in May 2026

Pure local (Tier 2+ hardware)

• ★ ★ ollama/qwen2.5-coder:32b — the canonical local Aider model. Q4_K_M handles diff format reliably.
• ★ ollama/qwen3:32b-instruct — strong general; good in architect role.
• ★ ollama/deepseek-r1-distill-llama:70b — architect mode; strong reasoning.

Hybrid local + API

• Architect: openrouter/anthropic/claude-sonnet-4.7 (or claude-haiku for cheaper).
• Editor: ollama/qwen2.5-coder:32b.
• Hard tasks routed to API; routine work stays local.

Pure API

• ★ ★ Claude Sonnet 4.7 (April 2025) — strong all-rounder.
• ★ Claude Opus 4.7 — for the hardest tasks.
• GPT-5 — competitive.
• DeepSeek V3 / R1 via DeepSeek API — cheap, strong, but data goes to DeepSeek.

Repo map

Aider parses your code with tree-sitter and sends a compact "this is what's defined" skeleton to the model. Configurable size:

aider --map-tokens 2048    # default
aider --map-tokens 8192    # bigger map for big repos
aider --map-tokens 0       # disable repo map entirely

The repo map is one of Aider's distinguishing features; it makes the model "aware" of what exists without sending every file.

.aider files

• .aider.conf.yml — per-project config (model, tests, lint).
• .aider.model.metadata.json — per-model overrides if needed.
• CONVENTIONS.md — Aider auto-loads this if present; great place to put coding conventions the model should follow.

Common honest gotchas

• Local models drop diff format sometimes. Q4 of a 7B will fail at this; Q4 of a 32B is usually fine; Q4 of a 70B is reliable. Use 32B+ for serious editing.
• Context window matters. Default Ollama tags are 4K–8K context; increase via Modelfile (PARAMETER num_ctx 32768) or per-call.
• Tests as guardrail. --test-cmd is the magic ingredient — the model sees failures and fixes. Use it.
• Auto-commit can pile up small commits. Use aider --no-auto-commits if you prefer to commit yourself.
• Long-running sessions. Aider remembers the chat; context grows; eventually you /clear and start fresh.
• /voice requires Whisper. Local Whisper or OpenAI; configurable.

Voice mode

• /voice toggles dictation.
• Uses whisper.cpp under the hood, or OpenAI's API.
• Surprisingly natural; "add error handling around the database call" works as well as typing.

Aider vs. Cline vs. Continue

• Aider — terminal, diff-based, repo-map, architect mode. ★ ★ for "I want a pair programmer that works well with local models."
• Cline — VS Code, more autonomous, the model writes whole multi-file changes and runs commands. Excellent with API, less reliable with local 30B.
• Continue.dev — VS Code, less autonomous, predictable; great FIM tab completion.

If you're a terminal-first developer, Aider. If you live in VS Code and want chat-side-panel, Cline or Continue.

Honest budget framing

• Pure local: zero marginal cost beyond electricity (~$10–30/month for a 3090 left running).
• Mixed: $10–50/month in OpenRouter / Anthropic credits at moderate use.
• Pure Claude Sonnet 4.7 API: $30–150/month for serious coding use.
• Claude Code subscription: $200/month for unlimited Sonnet/Opus access — for heavy users this beats per-token API pricing.

For more on this math see cost-cloud-vs-local.

Pick this if…

• One pick for terminal-based AI pair programming with local models: ★ ★ Aider + Qwen2.5-Coder 32B.
• Hard refactors / architectural decisions locally: Architect mode with R1-distill 70B.
• Hybrid: Aider routing to Claude Sonnet via OpenRouter for hard tasks, Qwen2.5-Coder 32B local for routine.
• Voice-driven coding: /voice mode with Whisper.
• You want IDE-side instead of terminal: Cline or Continue.dev.
• You want fully autonomous "go fix this": OpenHands — but plan for less reliability.