Launch HN: Codebuff (YC F24) – CLI tool that writes code for you

Hacker News Discussion ↗

A new CLI-based AI coding tool, Codebuff (formerly Manicode), promises to act as an autonomous “coding agent” that understands a local codebase, selects relevant files automatically, edits multiple files at once, and runs terminal commands and tests without constant user approval. Early adopters report significant productivity gains and praise its ergonomics compared with IDE-centric tools like Cursor, Aider, and Cline, especially for refactoring and working across monorepos. Skeptics question its $99/month pricing, reliance on proprietary cloud models, closed-source nature, and the security implications of allowing an LLM to run unrestricted shell commands and see project code, arguing that open-source or in-editor alternatives may offer better control and value.

Product & UX Overview

  • Codebuff is a CLI-based coding agent that can read a repo, choose relevant files, edit them, and run terminal commands/tests without manual file selection or per-command approvals.
  • It aims to act like a “junior engineer” or “skilled surgeon”: minimal diffs, multi-file edits, and iterative test/fix loops.
  • Designed to live beside any editor (VS Code, JetBrains, Neovim, Zed, etc.) in a terminal split rather than being an IDE plugin.

Comparisons to Other Tools

  • Repeated comparisons to Aider, Cursor, Cline, Cody, Amazon Q, etc.
  • Supporters highlight:
    • Auto file selection, deeper context, and single-shot multi-file edits as major UX wins.
    • True agent behavior (write tests → run tests → fix errors → rerun).
  • Critics argue:
    • Aider and Cline already offer similar capabilities (repo maps, treesitter, command execution, auto-approve modes).
    • Some prefer explicit file selection for safety/cost control.
    • For many, IDE-based tools are more convenient than a separate CLI.

Context & Technical Approach

  • Uses large-context models (mainly Claude 3.5 Sonnet) plus a preprocessing pass that scans the repo (file tree, function/class names) to ask a smaller model which files to read.
  • Team initially thought this was “not RAG” but discussion converges that any search-then-augment flow is a form of RAG.
  • Supports language-aware parsing via treesitter; some languages (e.g., Svelte) only partially supported.
  • Earlier approach used patch generation with custom apply logic; later changed due to reliability issues.
  • Encourages knowledge.md files to encode project-specific conventions and style guides.

Real-World Use, Strengths, and Weaknesses

  • Several users report strong productivity gains on real projects (Go/TS/Terraform monorepos, Elixir, Rust, Node/TS, Flutter, Python web apps).
  • Especially praised for refactors, test-writing, and multi-file changes; less compelling for tiny, precise edits where IDE tools are faster.
  • Some reports of incorrect or incomplete edits (e.g., overwritten modules, missed subclasses), but usually caught via diffs/CI.

Pricing & Credits

  • Pricing is ~$99/month with a credit system; excess usage bills per credit.
  • Many commenters view this as expensive relative to Cursor, Cody, and roll-your-own API usage.
  • One user’s $500 usage stemmed from a bug that granted excessive credits; this raised concerns about runaway costs and desire for hard limits and per-request cost visibility.

Security, Privacy, and Closed-Source Concerns

  • Codebuff can run arbitrary shell commands without explicit confirmation, which alarms some users.
  • Team argues:
    • In practice this has not caused serious issues.
    • Git plus an internal undo can recover from destructive actions.
    • Models are prompted to be cautious; directory resets try to keep commands scoped to the project.
  • Critics worry about:
    • Potential exfiltration of SSH keys, secrets, or personal data.
    • Accidental system-wide changes (e.g., Python installs, global packages).
    • Lack of sandboxing/VMs and reliance on “trust the model.”
  • Suggestions include sandboxing (VMs, pledge-like mechanisms, Docker), optional approval prompts, and better guardrails for untrusted repos.
  • Hosting is via the vendor’s servers, forwarding to LLM APIs; no bring-your-own-key option today. Some dislike the closed-source, cloud-only model and prefer local or self-hosted solutions.

Positioning, Differentiation, and Skepticism

  • Supporters say its simplicity, no-click workflow, and aggressive context gathering make it feel qualitatively better than other agents, especially in messy or mid-size codebases.
  • Skeptics see “just another wrapper” around third-party models, with features already present in mature open-source tools at lower cost.
  • Some question long-term viability without clearer differentiation, stronger privacy guarantees, or open-sourcing.
  • There is debate over CLI vs IDE as the primary interface: some love the terminal-first design; others see it as friction compared to embedded IDE assistants.

Feature Requests & Future Directions

  • Requests include: multi-repo support beyond a single directory, better handling of giant files, sandboxed execution, local/self-hosted models, improved docs and demos on large/codebase work, and benchmarks like SWE-bench.
  • Team mentions plans for privacy modes, possible sandboxing, and more complex demos (including dogfooding on their own production code).