dalao.ai

Aider

AI pair programming in your terminalOpen source

About Aider

No commits to the main repository since 2026-05-22 Commit history

Aider is an open-source AI pair programmer for the terminal: the user describes a change in a Git repo, and the user's chosen model edits the files while aider commits each change. It is free under Apache-2.0; model usage is billed by the provider. The latest GitHub release is v0.86.0, published on 2025-08-10; newer model settings and Python 3.14 support on main are unreleased. Main's last commit was on 2026-05-22; as of 2026-10-25 it had 49,744 stars, 1,407 open issues and 533 open PRs. Two CVEs made public on 2026-05-31 had no fix as of 2026-10-25.

EditorialUpdated 13 min read

The Aider interface

What it is

Aider is a command-line program. The user starts it in a Git repository, adds the files to change and describes the change in plain language; it sends the request plus the relevant code to an LLM, writes the model's edits back to disk and commits them to Git with a descriptive message. The repo's tagline is "aider is AI pair programming in your terminal", and the README says it "lets you pair program with LLMs to start a new project or build on your existing codebase."

No model comes bundled. Users bring their own provider API keys; per the official docs, it works with cloud models from Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter and others, and with local models through Ollama or any OpenAI-compatible endpoint. The code is mostly Python and ships on PyPI as aider-chat.

Who runs it: the aider.chat FAQ says Aider AI LLC is the company behind aider, and the privacy policy is issued by that company. Paul Gauthier made the first commit.

Public numbers (GitHub REST and Search APIs, as of 2026-10-25): 49,744 stars, 5,104 forks, 1,407 open issues and 533 open PRs. The newest commit on main, 5dc9490, is a PR merge at 22:02 (UTC+8) on 2026-05-22. README badges claim 6.8 million cumulative installs and 15 billion tokens a week; both figures are the project's own.

Aider lets you pair program with LLMs to start a new project or build on your existing codebase.

Aider's official README

Core capabilities

Repo map. This is the core of what separates aider from tools that paste whole files into a prompt. The docs explain that every request carries a compact map of the whole repository listing the most important classes and functions in each file with their signatures, so the model can see how a change relates to the rest of the code without reading all of it. The map is built by parsing source with tree-sitter and trimmed by importance to fit a token budget. Bash parsing rules were merged on main on 2026-05-16 and are not in a release yet.

Deep Git integration. Per the Git docs: started in a folder without a repo, it offers to create one; it commits after every edit, with a message written by a cheaper "weak model" from the diff and chat, following Conventional Commits by default; and if a file it is about to edit has uncommitted work, it commits that first so the user's changes and the AI's stay apart. /undo reverts the last AI change, /diff shows changes, and /commit and /git work on the repo directly. Commits written by aider get "(aider)" appended to the author name.

Four chat modes. The default code mode edits files directly; ask mode discusses without editing; architect mode splits the work between two models, an "architect" that proposes a solution and an "editor" that turns it into concrete file edits; help mode answers questions about aider itself. The docs suggest settling the plan in ask mode, then switching to code mode and saying "go ahead."

Use from any editor. Launched with --watch-files, aider watches the repo: when the user writes a one-line comment ending in AI! in any editor, it edits code to match; when the comment ends with AI?, it answers a question. No dedicated editor plugin is needed.

Other features. The README also lists support for 100+ programming languages, adding images and web pages to the chat, voice input, automatic linting and testing after each edit with attempted fixes, and copying context into web chat tools and pasting the answer back.

Benchmarks. The project maintains the aider leaderboard, which uses 225 Exercism exercises (C++, Go, Java, JavaScript, Python, Rust) to test how well each model edits code as asked inside aider. In the leaderboard data file, the top score belongs to gpt-5 (high) at 88.0% correct, costing $29.08 for a full run, dated 2025-08-23. Two caveats apply: these are model scores inside aider's harness, not a comparison of aider with other coding tools; and the data file was last updated on 2025-10-04, so models configured on main since then, such as Claude Opus 4.5 to 4.7 and GPT-5.5, are not on the board.

Who it's for / Not for

Good fit:

  • Developers comfortable in a terminal and Git who want every AI edit as its own commit they can undo at any time.
  • People who want to choose their own model and control model spend, including those running local models through Ollama to keep code off cloud services.
  • Team members on different editors: watch mode is triggered by code comments, not a plugin for one editor.
  • Anyone comparing how well models edit code: the leaderboard publishes its method and data.

Poor fit:

  • People who want a GUI, a chat sidebar and inline diff review.
  • Anyone expecting the tool to plan multi-step tasks, run commands on its own and work unattended for long stretches: aider works mostly one request at a time, and the user starts every turn.
  • Teams that rely on frequent releases and timely security fixes: the latest GitHub release is still v0.86.0 from 2025-08-10, two public CVEs have no fixed version, and as of 2026-10-25 main's latest commit was on 2026-05-22.
  • Environments limited to Python 3.13 or newer: v0.86.0 requires Python 3.10 to 3.12, and experimental 3.14 support exists only on main.

How it compares

Versus Claude Code. Claude Code is Anthropic's coding tool; Anthropic's pricing page (as of 2026-10-25) says it is included in all paid plans, with Pro at $20 a month billed monthly or $17 a month billed annually ($200 up front), and terminal use shares one usage allowance with chat. It is tied to Anthropic's models; aider can switch to any model but leaves API keys and spend to the user.

Versus OpenAI Codex CLI. Codex CLI also runs in the terminal, describes itself as a "lightweight coding agent that runs in your terminal", is released under Apache-2.0 and had 130,839 stars as of 2026-10-25. Its README recommends signing in with a ChatGPT account to use it as part of a Plus, Pro, Business, Edu or Enterprise plan, with an API key as the alternative.

Versus OpenCode. OpenCode describes itself as "the open source coding agent", is MIT-licensed, had 221,304 stars as of 2026-10-25, and was last pushed to that same day. Like aider, it is not tied to one model vendor.

Versus the community fork cecli. cecli, formerly aider-ce, is marked on GitHub as a fork of Aider-AI/aider; the repo was created on 2025-08-03, is released under Apache-2.0, had 415 stars as of 2026-10-25 and was still being pushed to that day; its latest release, v1.9.0, came out at 21:28 (UTC+8) on 2026-10-17. On 2025-11-13 a user wrote in aider's issue #4639 that the focus of community development had moved to this fork.

Where the prices come from: Claude Code's prices are from Anthropic's pricing page; for Codex CLI, the README states that it can be used with ChatGPT plans; OpenCode and cecli, like aider, are free under open-source licences with model costs on top. As of 2026-10-26, aider's latest GitHub release is v0.86.0, published at 01:42 (UTC+8) on August 10, 2025. Over the same stretch, Codex CLI's latest release, rust-v0.164.0, came out at 12:48 (UTC+8) on October 23, 2026; OpenCode's v1.19.4 at 21:03 (UTC+8) on October 22; and cecli's v1.9.0 at 21:28 (UTC+8) on October 17.

Getting started

Install. The recommended route is aider-install: the steps are python -m pip install aider-install, then aider-install. It puts aider in its own Python environment and installs a separate Python 3.12 if needed. Mac and Linux users can also run the one-liner curl -LsSf https://aider.chat/install.sh | sh, and Windows users the matching PowerShell command; these one-line scripts are built on uv's installer. uv tool install or pipx install aider-chat work as well. The docs warn that system package managers often install aider with incorrect dependencies.

Version requirements differ by source. v0.86.0 requires Python 3.10 to 3.12. The install page says the pipx and pip routes support 3.9 to 3.12, which does not match v0.86.0's requirement. Commit 861a1e4, merged on main at 06:00 (UTC+8) on 2026-03-10, raised the ceiling to 3.14 (experimental), and that change has not been released.

First run. In the project folder, the user sets the model provider's API key and runs aider followed by the files to edit. Common switches: --model sets the model; --architect starts in architect mode; --watch-files turns on comment triggers; --no-auto-commits turns off automatic commits. Every option is in the options reference and can also go in a config file or environment variables.

Each aider edit is a separate commit, so /undo reverts it and Git can roll back the whole branch; on a separate branch, those rollbacks leave main untouched.

Pricing & license

The software is free. aider is released under Apache-2.0, at no charge for personal or commercial use, and the code has no licence checks or paid features. The real cost is model API usage at the rates of the provider in use; aider shows an estimated cost for each message and for the whole session after each exchange. The leaderboard offers a reference point: the same 225 exercises cost about $29.08 per run with gpt-5 (high) and about $0.88 with DeepSeek-V3.2-Exp (Chat) (dated 2025-08-23 and 2025-10-03). Local models carry no API cost, but edit quality depends on the model.

Licence history. The repo's first commit, at 10:05 (UTC+8) on 2023-04-04, had no licence file. The licence was first added by Paul Gauthier's commit 34fd191 ("Added Apache 2.0 license"), committed at 23:29 (UTC+8) on 2023-05-15, which created LICENSE.txt with the full Apache-2.0 text. The GitHub repo was created at 02:57 (UTC+8) on 2023-05-10. That file remains the repo's only licence file, with the copyright-owner line never filled in; the pyproject.toml classifier is also Apache Software License. The FAQ likewise states that aider is open source under Apache 2.0. Third-party dependencies such as litellm and tree-sitter carry their own licences.

Risks & limits

Prompt injection via architect mode (CVE-2026-10175). At 22:26 (UTC+8) on 2026-04-21, a user reported in issue #5058 that in architect mode the architect model's output is handed to the editor model as is (editor_coder.run with preproc=False), bypassing the preprocessing applied to normal user input. The reporter hid malicious instructions in a repo README; the architect model followed them and wrote a line of code that sends out the contents of .env, the editor model wrote it to a file, and aider then committed it. Preconditions: the user adds attacker-controllable content (such as someone else's README) to the chat and uses architect mode; the reporter's reproduction used gpt-4o-mini, with --yes, on main's 0.86.3.dev19. In the code, --auto-accept-architect is on by default, so "Edit the files?" is not asked by default; --auto-commits is also on by default. VulDB, as the CNA, published the record at 16:45 (UTC+8) on 2026-05-31 and scored it 6.3 under CVSS 3.1 and 5.3 under CVSS 4.0; NVD's page shows the same source's 4.0 metrics as 2.1 (Low), NVD has no score of its own, and the record's status is Deferred. CISA-ADP's SSVC assessment: no exploitation seen, not automatable, partial technical impact. The "auth.py" in the description is the file injected during the reporter's test, not aider code; "0.86.3" is a development version number that was never released. As of 2026-10-25 the issue is still open with no maintainer reply, and the relevant code is unchanged in both the release and main.

No private-network guard on web scraping (CVE-2026-10177). At 01:08 (UTC+8) on 2026-04-27, the same reporter pointed out in issue #5075 that aider's built-in scraper does not check, before fetching, whether the target is a private address or a cloud server's metadata address. In the code, scraping only applies to URLs in messages the user sends, or to the /web command: when a URL is detected it asks "Add URL to the chat?" by default, --detect-urls is on by default, and with --yes-always the question is skipped. Exploitation requires such an address to appear in the user's message and be confirmed (or auto-confirmed), with aider running on a machine that can reach internal or metadata services. VulDB published the record at 18:15 (UTC+8) on 2026-05-31 with the same scores as CVE-2026-10175; CISA-ADP assesses a proof of concept as existing. The "api_docs.py" in the record is a file the model generated during testing. Fix PR #5137 was opened at 09:18 (UTC+8) on 2026-05-16 and remains unmerged as of 2026-10-25.

Git hooks skipped by default. Per the Git docs and options reference, aider commits with --no-verify by default, so pre-commit hooks do not run; hooks run only when --git-commit-verify is set.

Usage analytics. The official docs say analytics are collected only with the user's consent, are anonymous and tied to a randomly generated UUID, and never include code, chats, keys or personal information. When no flag is set, a random subset of users is asked whether to opt in, a share the code sets at 10%; data goes to PostHog. Running aider --analytics-disable once turns it off permanently. The privacy policy's last-updated date is 2025-04-12.

Release and maintenance pace (as of 2026-10-25). The newest entry on GitHub Releases is v0.86.0, published at 01:42 (UTC+8) on 2025-08-10. Since v0.86.0, main has added settings for Claude Opus 4.5 to 4.7, Sonnet 4.6 and GPT-5.3 to 5.5, a /ok command, a symlink-loop fix and experimental Python 3.14 support, none of it released; v0.86.0 users who want those models must configure them by hand or install from main. Per the GitHub API, as of 2026-10-26 the latest commit on main is the merge of PR #5173 (an Anthropic model-list expansion, commit 5dc9490) at 22:02 (UTC+8) on 2026-05-22, and the newest GitHub release is still v0.86.0. The README's list of models aider "works best with" still names older ones such as Claude 3.7 Sonnet, DeepSeek R1 and OpenAI o1.

How it got here

Beginnings. Paul Gauthier made the repo's first commit ("initial") at 10:05 (UTC+8) on 2023-04-04; the GitHub repo was created on 2023-05-10, the Apache-2.0 licence was added on 2023-05-15, and the first GitHub release, v0.5.0, was published at 07:29 (UTC+8) on 2023-06-08. The repo map began in May 2023 as a ctags-based approach, was renamed repomap around 2023-05-25, and later moved to tree-sitter. From 2023 the project blog kept publishing model evaluations, which grew into the leaderboard. The Aider-AI GitHub organisation, created on 2024-06-09, later became the repo's home.

2024 to early 2025. Over this period came the architect/editor two-model setup in v0.58.0 (2024-09-29) and --watch-files in v0.67.0 (2024-12-06). Each release note says "Aider wrote xx% of the code in this release"; per the FAQ, this is the share of new lines attributed to aider by a git blame-style count, and v0.86.0 reports 88%.

From June 2025. v0.86.0 was published on GitHub on 2025-08-10, with no newer GitHub release since. At 07:59 (UTC+8) on 2025-11-13, a user wrote in issue #4639 that the author was "currently working on other things", asked for successor maintainers to be named or for a note in the README, and suggested pointing users to the aider-ce fork; Paul Gauthier then closed the issue as not planned, without comment. From February to May 2026, main still took new model settings and community PRs, the last being an Anthropic model-list expansion merged on 2026-05-22.

Our verdict

aider turned AI coding in the terminal into a clear way of working: the repo map handles context, per-edit commits make changes traceable and reversible, the ask, code and architect modes cover everything from discussion to editing, and watch mode is not tied to any editor. It is not tied to a model and charges nothing for the software, so cheap or local models can be swapped in to suit the task. Its leaderboard method is public and its data inspectable, but the board measures models, not aider against other tools, and the data stops at 2025-10-04.

Three things need weighing. First, release and main have diverged: new model settings and Python 3.14 support added after v0.86.0 live only on main, and GitHub Releases still shows v0.86.0. Second, two CVEs made public in 2026 have no fixed release, and the scraping fix PR has gone unmerged since 2026-05-16; with architect auto-accept, auto-commits and skipped Git hooks all on by default, the risk is higher when working with untrusted repos. Third, as of 2026-10-25 main's latest commit is still the 2026-05-22 merge, with 1,407 open issues and 533 open PRs.

By use case:

  • For solo work in a developer's own repos, aider commits each edit and /undo reverts the last one; v0.86.0 supports Python 3.10 to 3.12.
  • With other people's repos or web content, the risk sits in architect auto-accept, the --yes and --yes-always options, and Git hooks skipped by default (--git-commit-verify turns them back on).
  • On cloud servers, issue #5075 reports that web scraping doesn't block internal addresses or cloud metadata endpoints, and the fix PR is still unmerged.
  • Maintenance: as of 2026-10-25 the main repo's latest commit was on 2026-05-22, while the cecli fork shipped v1.9.0 at 2026-10-17 21:28 (UTC+8).

Why it stands out

Aider is what happens when one engineer decides the terminal is enough. There is no sidebar, no diff viewer, no agent theatre. You point it at a repository, describe the change, and it edits files and commits with a message that reads like a person wrote it.

The technical spine is the repo map: a tree-sitter derived, ranked summary of your codebase that fits in a few thousand tokens and lets a model reason about a project it has never fully seen.

It works from the terminal, and its repo map fits a ranked summary of the codebase into a few thousand tokens.

Sources

Official

  1. GitHub - Aider-AI/aider: aider is AI pair programming in your terminalGitHub · Accessed Oct 26, 2026
  2. aider/README.md at 5dc9490bb35f9729ef2c95d00a19ccd30c26339c · Aider-AI/aiderGitHub · Accessed Oct 26, 2026
  3. aider/HISTORY.md at 5dc9490bb35f9729ef2c95d00a19ccd30c26339c · Aider-AI/aiderGitHub · Accessed Oct 26, 2026
  4. Releases · Aider-AI/aiderGitHub · Accessed Oct 26, 2026
  5. Merge pull request #4899 from turian/codex/python-3-14-support · Aider-AI/aider@861a1e4GitHub · Accessed Oct 26, 2026
  6. Added Apache 2.0 license · Aider-AI/aider@34fd191GitHub · Accessed Oct 26, 2026
  7. initial · Aider-AI/aider@41aed41GitHub · Accessed Oct 26, 2026
  8. Installation | aideraider.chat · Accessed Oct 26, 2026

Updates & reading