Show HN: Plandex – an AI coding engine for complex tasks

Original Article ↗ Hacker News Discussion ↗

Plandex is an open-source AI coding tool that uses GPT‑4 and similar models to plan and apply multi-file code changes in a git-backed “sandbox,” aiming to handle larger, more complex tasks than inline assistants like GitHub Copilot. Commenters highlight its branching workflows, diff review TUI, and integration with existing git practices, while debating trade-offs such as reliance on OpenAI’s API, lack of richer model support (Claude, Gemini, local models) and the value of its paid cloud offering versus self-hosting. Many see promise in its UX for repo-level development and iterative planning, but note current limitations in reliability, testing integration, and truly complex software engineering tasks.

Overall Reaction

  • Many commenters find Plandex promising as an AI coding tool for larger, multi-step tasks.
  • Some are skeptical that it does anything substantively beyond “trivial web projects” or what ChatGPT already does with manual copy‑paste.
  • Several users report real productivity gains on non-trivial work (e.g., infrastructure, billing backends, SPAs), but still needing human oversight.

Positioning vs Other Tools (Copilot, aider, GitHub bots)

  • Compared to Copilot, Plandex is praised for:
    • Working across multiple files, not just around the cursor.
    • Having a persistent, branching “plan” with saved conversation states.
    • Showing diffs before applying changes and allowing selective acceptance.
  • Compared to aider:
    • Plandex emphasizes multi-step plans, version-controlled sandboxes, branches, and a diff-review TUI.
    • aider emphasizes tight native git integration, chat-style UX, and using existing git workflows for undo/branches.
  • Some prefer GitHub Issue → PR bots; others find that flow clunky and like Plandex’s tighter loop and CLI.

Architecture, Git, and UX

  • Plandex uses a client–server model with Go backend, PostgreSQL, filesystem, and git combined into a transactional system.
  • Plans are version-controlled separately from the repo, which:
    • Avoids mixing user and AI changes.
    • Works even in non-git directories.
  • Diff-review TUI gets mixed feedback:
    • Some like side-by-side review and sandboxing.
    • Others want plain git diff output or an “auto-apply” mode and to rely on their existing diff tools.

Models, Costs, and Context

  • Currently relies mainly on OpenAI (gpt-4-turbo); no fine-tuning yet.
  • Roadmap includes supporting Claude, Gemini, Mistral, open source models, and local APIs like OpenRouter/Together/ollama, with function-calling support as a key requirement.
  • Reported cost: “meaty” tasks for under ~$1; more complex or iterative tasks can reach $1–2.
  • Uses gradual summarization for long plans and aims to be less aggressive than ChatGPT’s opaque strategy.

Capabilities, Limitations, and Future Features

  • Handles multi-file edits, AWS infra, billing backends, CLI/HTTP handlers, TUIs, and bug-hunting, but still makes syntax and “noob-level” mistakes.
  • Reliability of file patching improved using chain-of-thought around line ranges, but still needs human review.
  • Lacks automated test-running and execution, though workflows using test output as context are suggested; automated checks are on the roadmap.
  • Plans to add multimodal (vision) models, more frontends (web UI, plugins), Azure/custom endpoints, and collaboration features.
  • Cloud offering exists alongside self-hosting; server adds state, DB, and future sharing features, which some question vs the subscription price.