Show HN: Plandex v2 – OSS AI coding agent w/ diff sandbox, full auto, 2M context (plandex.ai)

Hey HN! I’m Dane, the creator of Plandex (https://github.com/plandex-ai/plandex), an open source AI coding agent focused especially on tackling large tasks in real world software projects.

You can watch a 2 minute demo of Plandex in action here: https://www.youtube.com/watch?v=SFSu2vNmlLk

And here’s more of a tutorial style demo showing how Plandex can automatically debug a browser app: https://www.youtube.com/watch?v=VCegxOCAPq0

I launched Plandex v1 here on HN a little less than a year ago (https://news.ycombinator.com/item?id=39918500). It got a positive reception overall (plus some skepticism, which is fair enough), and a lot of early users and contributors. HNers helped to push it to the top repo on GitHub trending, where it remained for a few days, racking up over 8k stars.

Now I’m launching a major update, Plandex v2, which is the result of 8 months of heads down work, and is in effect a whole new project/product.

In short, Plandex is now a top-tier coding agent with fully autonomous capabilities.

It combines models from Anthropic, OpenAI, and Google to achieve better results, more reliable agent behavior, better cost efficiency, and better performance than is possible by using only a single provider’s models.

I believe it is now one of the best tools available for working on large tasks in real world codebases with AI. It has an effective context window of 2M tokens, and can index projects of 20M tokens and beyond using tree-sitter project maps (30+ languages are supported). It can effectively find relevant context in massive million-line projects like SQLite, Redis, and Git.

There are also some major UX improvements. Plandex v1 offered just a git-like CLI interface. You’d send prompts via the `plandex tell` command, either passing a string, a prompt file path, or popping up vim/nano.

While that CLI interface still exists, there’s now a much more intuitive chat-like REPL experience that gives you fuzzy autocomplete for commands and file paths. This is the new default/recommended way to use Plandex, though having the CLI interface available is also useful for scripting, or for piping in context/prompts from other commands.

A bit more on some of Plandex’s key features:

- Plandex has a built-in diff review sandbox that helps you get the benefits of AI without leaving behind a mess in your project. By default, all changes accumulate in the sandbox until you approve them. The sandbox is version-controlled. You can rewind it to any previous point, and you can also create branches to try out alternative approaches.

- It offers a ‘full auto mode’ that can complete large tasks autonomously end-to-end, including high level planning, context loading, detailed planning, implementation, command execution (for dependencies, builds, tests, etc.), and debugging.

- The autonomy level is highly configurable. You can move up and down the ladder of autonomy depending on the task, your comfort level, and how you weigh cost optimization vs. effort and results. The default autonomy level is ‘semi-auto’, which loads context automatically, but requires your approval before applying changes to your project and executing commands.

- Models and model settings are also very configurable. There are built-in models and model packs for different use cases, including `daily-driver` (the default pack), `reasoning` (to use Claude 3.7 thinking in place of plain 3.7), `strong` (for the toughest tasks where extra cost and latency are acceptable), `cheap` (for smaller tasks), and `oss` (for an all open source pack, which currently leans strongly on DeepSeek R1/V3), and `gemini-exp` for a pack ...