● Conductor users: the cloud service shuts down November 20, 2026. Back up and switch to CoderBaba →
#ai-coding-tools · 2 posts
News, guides and comparisons of AI coding tools, from autocomplete and IDE assistants to terminal agents and parallel-agent orchestrators, with open-source picks.
Meta has acquired AI coding startup Conductor, and its cloud service shuts down November 20. What happened, what changes and how to protect your work.
A step-by-step guide to running Qwen2.5-Coder, Devstral or DeepSeek-Coder-V2 on your own machine with Ollama and connecting it to your editor.
"AI coding tool" covers a wide range of products, from a plugin that finishes your line to an agent that opens a pull request on its own. Knowing which category a tool belongs to makes it much easier to compare options and to see what you would lose or gain by switching.
The oldest and most widespread category. As you type, the tool predicts the next few tokens or lines. Speed matters more than raw intelligence here: suggestions need to appear in a few hundred milliseconds. These tools use small models with fill-in-the-middle (FIM) support, which can see the code both before and after your cursor.
Open-source options: Continue and Tabby, paired with a small local model such as Qwen2.5-Coder 1.5B or CodeGemma 2B.
A side panel where you ask questions, paste errors and request changes to selected code. Quality depends on the model and on context retrieval, meaning how well the tool finds the right files to include in the prompt.
Open-source options: Continue and Cline, with a mid-size model such as Qwen2.5-Coder 14B or 32B.
The assistant can now act: create and edit files, run terminal commands and read the output, with you approving each step. This is where most developers first meet coding agents.
Open-source option: Cline, a VS Code extension that plans, edits and runs commands with approval at each step.
Command-line tools that work in your repository directly. They fit naturally into existing workflows, scripts and remote servers, and they are easy to run several at once.
Open-source options: Aider, which commits each change to git, OpenHands, an autonomous agent with a sandboxed runtime, and Qwen Code, tuned for Qwen3-Coder.
The newest category runs many agents at once, each in its own isolated workspace, and gives you a dashboard to review their work. Conductor was a prominent example before it was acquired by Meta. You can build the same pattern from open-source parts with git worktrees, as our Conductor alternative guide shows, and it's built into CoderBaba, our free, open-source command-line agent.
Tools that comment on pull requests, flag bugs and suggest fixes. They run in CI rather than in your editor, so they have to be precise and keep false positives low.
Before adopting a tool, especially for a team, check these seven points.
Closed tools are often polished and quick to start with. Open-source tools ask a little more setup in exchange for three lasting advantages:
Many developers end up with a hybrid: a polished closed tool where it shines, plus an open-source, local-capable fallback for sensitive code, offline work and cost control.
| Tool category | What the model needs | Good open models |
|---|---|---|
| Autocomplete | Small, fast, FIM support | Qwen2.5-Coder 1.5B or 7B, CodeGemma 2B, StarCoder2 3B or 7B |
| Chat | Instruction-tuned, good reasoning | Qwen2.5-Coder 14B or 32B, DeepSeek-Coder-V2 Lite |
| Agents | Tool use, long context, multi-step reliability | Devstral, Qwen3-Coder |
Browse full specs in the model directory.
Posts tagged #ai-coding-tools cover launches and major updates, acquisitions and shutdowns, hands-on comparisons and setup guides for connecting tools to open models. They are listed above, newest first.
Not necessarily. Some tools span several categories. Continue handles both autocomplete and chat, and Cline combines chat with agentic editing. Many developers still prefer a small, dedicated autocomplete setup plus a separate agent for bigger tasks, because the two have very different speed and model needs.
On the tooling side, open-source projects are competitive and often more flexible. Quality mostly comes down to the model you connect. With a strong hosted model, an open-source agent can do anything a closed one can. With a local model, results depend on its size and training. Agentic models like Devstral and Qwen3-Coder narrowed that gap considerably.
It can be, with sensible guardrails: work on a branch or worktree, keep command approval turned on, review every diff and run fully autonomous agents in a container or sandbox. Treat an agent like a new teammate with shell access. Be helpful, but verify.
That depends on where your state lives. If your tasks, branches and instructions are in git, switching tools is an afternoon's work. If they live only in a vendor's cloud, you are on the vendor's timeline, as Conductor users found with a two-week cloud shutdown notice. Keeping state in your repo is the best insurance.
Yes. Hands-on comparisons and setup guides are published under this topic. We disclose that we make our own agent, CoderBaba, and we aim to keep reviews fair and factual.