Aider vs Claude Code: Git-Native Pair or Full Agent

Aider and Claude Code both live in your terminal and both edit your repository from plain-English requests. That is where the similarity ends. The aider vs Claude Code choice is really two separate choices bundled together: how much of the loop you want to run yourself, and who gets to pick the model.

Aider is a pair programmer. You choose the model, you choose which files it can edit, and every edit it makes lands as a git commit you can inspect or throw away. Claude Code is a fuller agent harness. It explores the repo on its own, delegates to subagents, runs hooks and skills, connects to MCP servers, and works across terminal, IDE, desktop, web, and CI. In exchange, it only runs Claude models.

This post goes head-to-head on the parts that decide it. For the wider field of terminal agents, see Claude Code alternatives. For how Claude Code compares with the other closed-model agent, see Claude Code vs Codex.

Quick comparison

The real difference: who picks the model

Aider treats the model as a setting. Its LLM documentation says it can connect to most LLMs, and it has pages for providers including OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Amazon Bedrock, Vertex AI, and GitHub Copilot. Anything else goes through LiteLLM, which its docs say reaches hundreds of other models.

Local models work too. Its Ollama page walks through pulling a model and starting aider against it:

aider --model ollama_chat/<model>

There is a catch. Aider's docs warn that it may not work well with less capable models, because the model has to return edits in a format aider can apply. If a model returns code but nothing lands on disk, that is usually why.

Claude Code is built around Claude. You can reach those models through Anthropic's own plans or through Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry, as the model configuration docs describe. But Anthropic's gateway documentation is blunt: it doesn't support routing Claude Code to non-Claude models through any gateway.

So if you want to swap a cheaper model in for routine edits, try a different lab's latest release, or keep code on a model running in your own building, Aider does that natively. Claude Code does not.

Git: every edit a commit, or commits on request

This is Aider's signature, and it shapes the whole workflow.

According to Aider's git documentation, whenever it edits a file it commits the change with a descriptive message. If a file already has uncommitted changes, it first commits those separately, so your work and its work never mix. Commit messages follow Conventional Commits by default, and aider-authored commits get "(aider)" appended to the git author name.

That gives you three cheap controls. /diff shows what changed since your last message. /undo undoes the last commit if aider made it. And because every step is a commit, you can review a session with git log or run it on a branch.

Claude Code treats git as one more job you hand it. Its overview says it works directly with git: it stages changes, writes commit messages, creates branches, and opens pull requests, typically when you ask. For undo it has checkpointing: each prompt that starts a turn creates a checkpoint, and /rewind can restore the code, the conversation, or both.

One limit matters here. The checkpointing docs state that it does not track files modified by Bash commands. Aider's undo works on commits, so anything it committed can be reverted with plain git. It only covers what aider itself committed, though.

If you want a clean, reviewable trail without thinking about it, Aider's default is better. If you'd rather commit at a point you choose, that is fine with Claude Code, but you own that discipline.

How each one finds the right code

Aider builds a map. Its repo map documentation says it sends a concise map of your whole git repository, with key classes, functions, and signatures, alongside each request. A graph-ranking step picks the most relevant parts to fit a token budget, which defaults to 1k tokens and expands when no files are in the chat.

You also decide which files are editable. Aider's usage docs say you add files to the chat so it can edit them, and recommend adding only the files that need changes. It pulls in related context automatically, and the model can ask to see more files, which aider offers to add.

Claude Code explores. It reads files and searches on its own. Its permissions documentation lists file reads and Grep as read-only tools it can use without asking inside your working directory. When research would flood the main conversation, a subagent can do it in a separate context window and return a summary.

The tradeoff is predictability versus reach. Aider's behaviour is easy to reason about because you set the boundaries. Claude Code can follow a thread across the repo without being told where to look, which helps on unfamiliar code and costs more tokens.

Autonomy and permissions

Aider's modes are small and explicit. The chat modes docs list four: code, ask, architect (one model proposes, another turns the proposal into edits), and help. The recommended habit is to bounce between /ask to agree on a plan and /code to carry it out.

It can verify its own work. Aider's docs say it can lint and test your code every time it makes changes and try to fix what the linters or tests report. /run and /test run shell commands from inside the chat, and --message lets you script a single instruction from the command line. Its watch mode lets you stay in your own editor: add a comment ending in AI! and aider picks it up and makes the change.

# Add error handling for negative numbers. AI!

Claude Code's permission modes run a wider range, from one that reviews every action yourself, through accept-edits and plan, to an auto mode where a second model reviews actions instead of you, down to modes for locked-down CI and for isolated containers only. Its sandbox restricts what shell commands can reach at the operating-system level, though the docs note it is off by default and covers shell commands only.

For unattended work, claude -p runs non-interactively for scripts and CI, and there is an official GitHub Actions integration.

The pattern: Aider keeps you in the loop by design. Claude Code lets you dial the loop open or shut.

What Claude Code has that Aider's docs don't

I searched Aider's documentation site on 5 October 2026 for MCP, subagents, and skills, and found none of them documented. Its only hook-related mention is an option to run your git pre-commit hooks. Claude Code documents all four:

• Subagents with their own context window, tool access, and permissions.
• Skills, reusable procedures that load only when used, following the Agent Skills open standard.
• Hooks that run at lifecycle points and can block a tool call.
• MCP servers for tools like issue trackers, databases, and design files.

It also spans more surfaces. Per its overview, the same engine runs in the terminal, VS Code, JetBrains, a desktop app, and the web, and your CLAUDE.md files, settings, and MCP servers carry across them. Aider's counterpart is the editor-comment workflow above, plus a browser UI it describes as experimental.

If your workflow leans on any of those four, that settles it. If it does not, you may be paying for machinery you never use.

Is Aider still active?

Check before you standardise on it. As of 5 October 2026, the latest release on PyPI is 0.86.2, uploaded on 12 February 2026. The GitHub repository shows commits as recent as May 2026, and the release notes carry a section of changes on the main branch that haven't shipped in a release.

Its LLM page still names Gemini 2.5 Pro, DeepSeek R1 and V3, Claude 3.7 Sonnet, and OpenAI o3, o4-mini, and GPT-4.1 as its best models. Those are older models. Aider can still connect to newer ones, but expect to do more of the model plumbing yourself.

If that pace bothers you, the other open-source route is covered in OpenCode vs Claude Code.

Aider or Claude Code: how to pick

Many people run both. Aider for quick, supervised, commit-by-commit edits; Claude Code for the larger task you'd rather delegate. They don't conflict, because Aider's commits are plain git.

If neither one is your actual bottleneck

Both tools assume a working development environment and a repo you can run. For plenty of people the wall comes earlier: they watch someone share an AI setup that does exactly what they need, and never get it running.

Taku is aimed at that gap. It is an AI-native desktop workspace where you mirror an AI app, skill, or workflow someone already got working, run it, remix it, and keep it. It isn't a coding agent and doesn't compete with Aider or Claude Code. Taku is in Beta, and the Mac app is available now.

FAQ

Is Aider free?

The software is open source under Apache-2.0. But it calls model APIs, and you pay each provider for what you use, unless you run a local model. Claude Code's billing is tied to a Claude subscription or Anthropic Console account instead. Check each vendor's current pricing page rather than trusting a comparison.

Can Aider use Claude models?

Yes. Aider's Anthropic page has you supply an Anthropic API key and run it against the Claude models Anthropic offers.

Can Claude Code use other models?

Not in a way Anthropic supports. It can run Claude models through Anthropic, Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry, and Anthropic's docs say it doesn't support routing to non-Claude models through a gateway.

Which one is safer to let loose on a repo?

Aider's one-commit-per-edit habit makes every change easy to review and revert. Claude Code offers more guardrails you can configure, including permission modes and an OS-level sandbox, but they are options you turn on. Neither should run unattended against production without deliberate limits.

Does Aider support MCP, subagents, or skills?

Not according to its documentation, which I searched on 5 October 2026. Claude Code documents all three, plus hooks.

Can I use Aider and Claude Code on the same repository?

Yes. Aider's commits are ordinary git commits, and Claude Code works from your working tree, so they share a repo cleanly. Commit or stash first so Aider's dirty-file commit doesn't sweep up work in progress.