Aider

Aider is a chat loop, not an autonomous agent. You add specific files to the session, describe a change, and the model returns edits in one of several structured formats (search/replace blocks, unified diffs, whole-file rewrites — the repo carries a separate coder implementation for each, in aider/coders/). Aider applies them and commits. It has an /architect mode that splits planning and editing between two models, an /ask mode that answers without touching files, and a watch mode where you leave # ... AI! comments in your editor and aider picks them up and acts on them (https://aider.chat/docs/usage/watch.html).

What it does not have is a tool-calling agent loop or MCP. There is no MCP module anywhere in the aider/ package tree, and aider/args.py on main contains zero occurrences of the string mcp (https://api.github.com/repos/Aider-AI/aider/contents/aider, https://raw.githubusercontent.com/Aider-AI/aider/main/aider/args.py). Native MCP support has been requested in open issues and remains unimplemented (https://github.com/Aider-AI/aider/issues/4506). Third-party bridges exist but are community projects, not part of aider.

For some buyers that absence is the point: aider does what you asked to the files you named, and stops. For anyone expecting the tool to explore the repo, run commands, and iterate unsupervised the way current agent CLIs do, this is the wrong shape of product.

Aider builds a condensed map of the whole git repository — the important classes and functions with their signatures — and sends that instead of the codebase. It parses with tree-sitter (tree-sitter-language-pack==0.13.0 is a pinned dependency) and then ranks what to include with a graph algorithm over a dependency graph where files are nodes and imports/references are edges, so the map shows the identifiers most referenced by other code. The budget defaults to 1k tokens via --map-tokens and expands when no files are in the chat (https://aider.chat/docs/repomap.html, https://raw.githubusercontent.com/Aider-AI/aider/main/requirements.txt).

This is the part of aider worth respecting on its own terms. It is a cheap, deterministic answer to codebase context that predates and undercuts the brute-force approach of feeding a million-token window. If you are cost-sensitive, or running a local model with a small context window, this design does real work for you.

Every edit aider makes is committed with a generated message, and it commits any pre-existing dirty state first so its own changes are isolated. /undo discards the last change, /diff shows what changed since your last message, /git drops you to raw git. All of it can be turned off: --no-auto-commits, --no-dirty-commits, --no-git (https://aider.chat/docs/git.html).

Aider also lints every file it edits by default and will attempt to fix errors when the linter exits non-zero; point it at your test suite with --test-cmd plus --auto-test and it does the same for failing tests (https://aider.chat/docs/usage/lint-test.html). The default-on linting is unusual and sensible — it catches the syntactically broken edit before you do.

Aider is Apache-2.0, has no hosted tier, no pricing page, and no company behind it — the homepage links only to docs, GitHub, PyPI and a Discord (https://api.github.com/repos/Aider-AI/aider, https://aider.chat/). Your entire spend is LLM API usage, and aider reports token cost per request.

Model routing goes through LiteLLM (litellm==1.82.3, pinned), and the docs cover 15 providers including Ollama and LM Studio for local models, OpenRouter, GitHub Copilot, Bedrock and Vertex (https://aider.chat/docs/llms.html). There is also a --copy-paste mode for people whose employer only permits a web chat UI, or who would rather use a flat-rate subscription than an API key: aider keeps your clipboard loaded with the current code context and detects the pasted reply. The docs are unusually candid that this is legally awkward — "Most LLM web chat TOS prohibit automating steps (2) and (3) where code is copied from and pasted into the web chat" — and aider leaves those two steps manual by design (https://aider.chat/docs/usage/copypaste.html). Read your provider's terms before you build a workflow on it.

This is the finding that should change a purchase decision, and it is all from platform-recorded metadata rather than anyone's opinion.

As of 13 August 2026: the most recent commit to main is dated 22 May 2026, and GitHub's participation stats show roughly 136 commits across the last 52 weeks with zero in the most recent 12 weeks (https://api.github.com/repos/Aider-AI/aider/commits, https://api.github.com/repos/Aider-AI/aider/stats/participation). On PyPI, aider-chat went from a weekly-to-fortnightly release cadence through mid-2025 to two releases in the past twelve months: 0.86.1 on 13 August 2025 and 0.86.2 on 12 February 2026, which is still the current version (https://pypi.org/pypi/aider-chat/json). Ten pull requests were merged in all of 2026, against 461 open PRs and 1,333 open issues (https://api.github.com/search/issues?q=repo:Aider-AI/aider+type:pr+is:merged+merged:%3E2026-01-01).

The repository is not archived and carries no deprecation notice, and the maintainer was still merging community PRs in May 2026. It has 48,164 stars and 4,835 forks. Nothing here says abandoned; it says one-maintainer project that has slowed a great deal. Treat it accordingly — pin a version, expect to carry your own patches, and do not plan around an upstream fix arriving.

Concretely, three things.

Model defaults lag. Aider ships per-model settings (edit format, reasoning-token handling, temperature quirks). Commits on main from April–May 2026 add settings for GPT-5.5 and newer Claude models, but those commits are not in any release — 0.86.2 predates them (https://aider.chat/HISTORY.html). Newer models still work: you point at a fully-qualified provider/model name and supply your own .aider.model.settings.yml and .aider.model.metadata.json, and the docs say you can "safely ignore aider's warning about unknown context window size and model costs" (https://aider.chat/docs/config/adv-model-settings.html). You just do the tuning the project used to do for you, or you install from the main branch instead of PyPI.

The docs are stale in ways that will mislead you. The homepage still recommends "Claude 3.7 Sonnet, DeepSeek R1 & Chat V3, OpenAI o1, o3-mini & GPT-4o" and still cites 44,000 GitHub stars against an actual 48,164 (https://aider.chat/).

The leaderboard is a historical document. See below.

Aider's polyglot benchmark is one of the better-constructed public coding evals: the 225 hardest of Exercism's 697 exercises across C++, Go, Java, JavaScript, Python and Rust, deliberately recalibrated so top models would land between roughly 5% and 50% rather than saturating (https://aider.chat/2024/12/21/polyglot.html). It measures whether a model can both solve the problem and emit an edit aider can apply, which is a more honest test of a coding harness than pass@k on isolated functions.

Its problem is the date. The highest score in the published dataset is 88.0% for gpt-5-high, benchmarked in August 2025, and the most recent entry of any kind is DeepSeek v3.2 on 3 October 2025 (https://raw.githubusercontent.com/Aider-AI/aider/main/aider/website/_data/polyglot_leaderboard.yml). No model released after that appears. If you land on the leaderboard from a search result, you are reading a snapshot of late 2025, not a ranking of what you would actually use today.

The homepage claims 6.8 million pip installations and 15 billion tokens processed weekly. Neither is independently checkable and both should be read as vendor figures.

What is checkable: aider-chat downloads on PyPI, excluding mirrors, ran 854k (March 2026), 856k (April), 942k (May), 903k (June) and 838k (July), with 797k in the last 30 days (https://pypistats.org/api/packages/aider-chat/overall, https://pypistats.org/api/packages/aider-chat/recent). PyPI download counts include CI and container rebuilds and overstate human users, but the shape is the useful part: demand has been flat, not collapsing, right through the period when development stopped. People are still running it. That is a reasonable signal that the current release is stable enough to keep using.

cecli-dev/cecli is a GitHub-recognised fork of Aider, created 2 August 2025, Apache-2.0, 393 stars, with commits as recent as today (https://api.github.com/repos/cecli-dev/cecli). It describes itself as "yet another cli agent crafted for extensibility and customization" and credits "Both Cecli and Aider main" contributors; its roadmap covers agent mode, sub-agents and hooks — the things upstream aider does not have (https://raw.githubusercontent.com/cecli-dev/cecli/main/README.md).

We have not evaluated cecli and it is a young project with a small following. It is worth knowing about if aider's architecture suits you but its cadence does not.

The released 0.86.2 declares requires-python >=3.10,<3.13 (https://pypi.org/pypi/aider-chat/json), so a plain pip install aider-chat fails on a current default Python. The official aider-install package works around this by having uv provision Python 3.12 in an isolated environment, and the one-line install scripts do the same (https://raw.githubusercontent.com/Aider-AI/aider-install/main/aider_install/main.py, https://aider.chat/docs/install.html). Use those rather than pip into your system Python.

Main branch has since widened the constraint to >=3.10,<3.15 and added 3.13/3.14 classifiers (https://raw.githubusercontent.com/Aider-AI/aider/main/pyproject.toml) — another improvement sitting unreleased.

Analytics are opt-in, not opt-out. When enabled they go to PostHog and cover which models and how many tokens, which edit formats, feature and command usage, and exceptions; the docs state aider "never collects your code, chat messages, keys or personal info", and you can point it at your own PostHog instance (https://aider.chat/docs/more/analytics.html). We have not audited aider/analytics.py against that statement, but the source is public and short if your compliance team wants to.

Searching for Aider coverage from 2026 returns a large volume of review-style pages carrying figures that contradict the primary record — star counts of 39k, 41.6k or 44k, claims of "releases roughly every two weeks", and a last-commit date of March 2026. The GitHub API gives 48,164 stars, a last commit of 22 May 2026, and two releases in twelve months. Where this listing and those pages disagree, we have cited the API and PyPI endpoints above so you can re-run them yourself.