Local AI Pair Programming in 2026: Which Tools Are Still Maintained

Search for an AI pair programmer that runs against your own local model and you will be pointed at the same tool almost every time. It has 49,000 stars, it is genuinely good, and it has not shipped a release in more than a year. Nothing on its page says so. This guide checks which terminal and editor AI assistants are actually still being developed, using repository activity rather than reputation, and which of them work with a local model.

Quick answer

• Aider is the one everyone recommends, and it has gone quiet — last release 9 August 2025, and no commits at all in the 90 days to 21 September 2026.
• Cline is the most active — around 100 commits in the last 90 days. Continue is maintained but slower, with only 3 commits in the same window despite a push this week. Both are Apache-2.0.
• goose has moved — the old block/goose path now redirects, which breaks scripts and older tutorials.
• Check before you adopt. A stale AI coding tool is worse than a stale library: it is pinned to the model APIs and prompt formats of the month it stopped.

The maintenance table

Every figure below was read from the GitHub API on 21 September 2026. Stars are a measure of past popularity; the columns that matter are the last two.

“Last push” is any commit to the repository; “latest release” is the most recent tagged release. A tool can be healthy with infrequent releases, or with a long gap between them — but a repository with neither recent commits nor a recent release is the pattern to avoid.

The Aider problem, stated carefully

Aider is not abandoned in any formal sense. The repository is not archived, it is not disabled, and — this is the part that matters — its README carries no notice of any kind. It still describes itself as “AI Pair Programming in Your Terminal” and still says it “can connect to almost any LLM, including local models”. Someone arriving today has no way to tell.

What the API shows is a clean break in cadence. Releases ran roughly monthly through 2025 — v0.82.0 in April, v0.83.0 and v0.84.0 in May, v0.85.0 in June, v0.86.0 on 9 August — and then stopped. There have been no commits in the last 90 days, and 1,883 issues are open.

That is not a reason to delete it if it works for you today. It is a reason not to build a workflow on it, because an AI coding tool is unusually exposed to drift: model names change, provider APIs change, and tool-calling formats change. A library that stops moving keeps working. An AI client that stops moving slowly stops being able to talk to anything current.

What is actually maintained

Cline is the most active of the set: 68,892 stars, a push the day before this was written, and tagged releases days apart. It runs as a VS Code extension and drives an agent loop that reads and edits files, runs commands and iterates.

Continue is the other safe pick at 35,962 stars, pushed the same day this was checked. Note the gap between its activity and its release tags — the last tagged release is v2.0.0-vscode from 19 June 2026 while commits continue, though at only 3 in the last 90 days. That combination is normal for a project shipping from a main branch; it is the reverse of Aider’s pattern, where the tags and the commits both stopped.

goose is active at 54,503 stars, but with a catch worth knowing: the repository has moved. The widely-linked block/goose path now redirects to aaif-goose/goose. Redirects are transparent in a browser and silently confusing in a script — the GitHub API returns a “Moved Permanently” object rather than the data you asked for, which is the same trap that bites people using the old ComfyUI repository path.

Roo Code has also moved, from RooVetGit/Roo-Code to RooCodeInc/Roo-Code, and its last push was 15 May 2026, with its last release the same day and no commits in the last 90 days — the same profile as Aider, on a shorter clock. Update, 10 October 2026: the repository has since been archived on GitHub, which makes it read-only. That is the clearest signal a project can give.

Pointing any of them at a local model

All of these are clients, as is OpenCode, a terminal coding agent with its own setup for local models. None of them runs a model. You need a server, and the practical choice for most people is Ollama exposing its OpenAI-compatible endpoint, which these tools can be configured against as if it were a hosted provider. Our Ollama server and Docker guide covers running it safely, and the comparison of local LLM runners covers the alternatives.

Which model to load is a separate question with a real answer — see the best local LLMs for coding for the models worth using by VRAM tier, and running Qwen Coder in VS Code for a worked setup.

If you are writing the integration yourself rather than using an extension, Ollama with LangChain in Python covers the chat and embedding classes directly.

The constraint nobody mentions: context

Agentic coding tools are the hungriest context consumers in local AI. They send file contents, directory listings, command output and their own reasoning back into the window on every turn. A model advertising 128K of context needs the memory to hold 128K, and that memory is not free — see what a context window costs in VRAM for the arithmetic, and size a specific setup with the VRAM calculator.

This is usually what makes local agentic coding feel worse than the cloud equivalent. LM Studio Bionic takes a different route, letting each session switch between a local model and a hosted open model. It is rarely the model’s ability; it is that the context was quietly truncated.

How to check this yourself

The whole audit above is two API calls per repository, and it is worth running before you adopt anything in a field moving this quickly:

curl -s https://api.github.com/repos/OWNER/REPO \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(d['pushed_at'], d['archived'], d['open_issues_count'])"

curl -s https://api.github.com/repos/OWNER/REPO/releases/latest \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(d.get('tag_name'), d.get('published_at'))"

If the repository has moved, the first call returns a short JSON object containing message instead of the fields you expected. That is the signal to find the canonical path before trusting anything else you read about it.

FAQ

Is Aider dead?

Not officially. The repository is not archived and carries no notice. But there were no commits in the 90 days to 21 September 2026 and the last release was 9 August 2025, after a roughly monthly cadence. Treat it as dormant.

What is the best maintained AI pair programmer?

On repository activity alone, Cline — 68,892 stars, Apache-2.0, releases days old. Continue is the other actively developed option.

Do these work with local models?

Yes. They are clients, not runtimes. Point them at an OpenAI-compatible endpoint such as the one Ollama serves, and they treat it like any other provider.

Why did the goose repository link stop working?

It moved from block/goose to aaif-goose/goose. Browsers follow the redirect silently; the GitHub API returns a “Moved Permanently” object instead of repository data, which breaks scripts.

Does stars count tell me anything?

Only about the past. Aider has 49,086 stars and has not shipped in over a year. Check the last push and the last release instead.

Sources

• GitHub API, read 21 September 2026: Aider-AI/aider, cline/cline, continuedev/continue — stars, push dates, release tags, licences and open-issue counts

Roo Code archive status added 10 October 2026. Updated 21 September 2026. Repository activity was read from the GitHub API on that date and will drift; re-run the commands above before relying on it. Nothing here is a benchmark.