Aider is the strongest open option for developers who want an AI coding agent without a subscription or a closed runtime. It lives in the terminal, commits every change with a clear message so the diff stays reviewable, and points at any model behind any API key, including local ones via Ollama. That flexibility is the moat: when a new frontier model ships, Aider can use it the same day. The cost is ergonomic. There is no IDE, no inline autocomplete, and you own the token bill. For CLI-native developers who value model choice and transparent git history over a polished GUI, it is solid.
- Model agnostic: works with any API or local model via Ollama
- Auto-commits every change with a clear message, fully reviewable
- Free and open source, you only pay for model tokens
- Repo map gives the model whole-codebase context on large repos
- Architect plus editor mode splits reasoning from precise edits
- No IDE or GUI, terminal only with no inline autocomplete
- You own the token bill, frontier models add up fast
- Edit reliability varies by model and edit format
- Smaller feature surface than funded commercial agents
Aider is an open-source AI coding agent that runs entirely in the terminal. It pairs with a large language model of your choosing, edits files in your repository, and commits each change to git with a sensible message. The model is yours to pick: Claude, GPT, Gemini, DeepSeek, or a local model, all behind your own API key.
Where it fits
Aider is built for developers who live in the command line. There is
no editor extension and no chat panel: you run aider in a repo, add
the files you want in scope, and describe the change. Because it
commits every edit, the work stays inside normal git tooling, so you
diff, undo, and review AI changes exactly as you would a human
contributor’s. Its repo map feeds the model a structured view of the
whole codebase, which keeps multi-file edits coherent on larger
projects. The defining trait is that Aider is model-agnostic and
bring-your-own-key: it never resells inference, so you point it at
whatever model fits the task and pay the provider directly.
Cost to adopt
The tool itself is free under Apache-2.0. The real cost is the model API behind it. Driving a frontier model like a top Claude or GPT tier on a large refactor can run real money per session, and architect mode doubles request volume by design. The counterweight is choice: route routine edits to a cheap open model such as DeepSeek, or run fully local through Ollama for zero marginal cost. That economic control is the core argument against flat-rate subscription agents. You trade a predictable monthly fee for the ability to match spend to the job, which favors heavy users on cheap models and light users who would otherwise pay for an idle seat.
How it compares
Claude Code, Anthropic’s terminal agent, deeper autonomy but locked to Claude models and Anthropic billing. Pick Aider when model choice and bring-your-own-key economics matter more than single-vendor polish.
Cursor, a full AI-native IDE with inline completion and a GUI. Pick Aider when you want a terminal workflow, transparent git history, and no editor lock-in.
Cline, an open VS Code agent extension, also model-agnostic but lives in the editor. Pick Aider for a CLI-first, git-centric flow outside any IDE.
What changed recently
Aider’s most recent stable release, v0.86.0 on 2025-08-09, added support for the GPT-5 family and Grok-4, alongside clearer feedback on the /clear and /undo commands. The prior runs (v0.83 through v0.85 across May and June 2025) layered in newer Gemini versions, Claude Sonnet and Opus updates, enhanced thinking-token handling, and a --add-gitignore-files flag. The architect plus editor mode remains a defining feature: it routes a request to a reasoning model to plan the change, then hands the plan to a separate editor model that applies precise diffs, which helps with models strong at reasoning but weaker at exact file edits. The maintainer’s polyglot leaderboard (225 Exercism exercises across six languages) continues to track edit accuracy and cost, with GPT-5 high topping it at 88.0%.
Sources
- Aider official site, aider.chat, accessed 2026-06-05
- Aider releases, GitHub, latest stable v0.86.0 (2025-08-09)
- Aider polyglot LLM leaderboard, aider.chat, accessed 2026-06-05
- Aider chat modes (architect/editor), aider.chat, accessed 2026-06-05