jumpy 🐸

Inspired by 99 : I, like Prime, wanted a tool to allow for me to still be in the driver seat if I am going to spawn an AI to edit my code.

The main difference between jumpy and other tools, like 99, are:

feature	jumpy	99	avante / sidekick	codecompanion inline	claude code / aider
Interaction model	Inline prompt → hunks in buffer	search / visual / vibe → qfix or inline replace	Sidebar chat → apply	Inline edit → accept/reject	CLI agent → full file writes
Review granularity	Per-hunk accept/reject	Per-result (search) or selection replace (visual)	Per-file or per-suggestion	Per-change	Post-hoc (git diff)
Context switch	None: you stay in buffer	Minimal; qfix for search	Sidebar opens	Minimal	Leave editor or split terminal
LLM output format	Search/replace blocks	CLI provider output (opencode, claude, etc.)	Full file / patch	Full file	Full file via temp
Token efficiency	High (only changed lines sent back)	Mode-dependent; visual is targeted	Lower (full context)	Lower	Lower
Scope	Single targeted edit	Selection edit, project search, agentic work	Multi-file refactors, chat	Single edit or chat	Whole-project agentic tasks

Some of the existing features are to be fleshed out, but the sole purpose is to know exactly what I am letting the LLM write into my code. I have no interest in letting it change everything, and only then can I go back and review every change. That being said, here are the features of jumpy:

• inline prompt: describe a change without leaving the buffer
• search/replace hunks: LLM returns only changed sections, not whole files
• in-buffer diff review: proposed edits shown inline with accept/reject before anything is written
• per-hunk control: accept, reject, or reprompt individual hunks
• hunk navigation: jump between proposed changes with ]h / [h
• quickfix integration: collect pending hunks across buffers into one list
• token-efficient: sends only the targeted edit context, not full-file rewrites
• multi-provider: openrouter, openai, anthropic
• @lsp context: pull workspace symbols into the prompt when needed
• zero context switch: no sidebar, no terminal agent, no leaving your file

So the philosophy here is: yes, I still want to handwrite my code AND use AI, but I prefer that I have full control over what the AI is writing. I don't want to have to wait until it finishes in order to review its changes.

Contributing

Very open to PRs, issues, etc.! There is no concrete set of guidelines right now. I am quite welcoming to pretty much anything [1]!

[1]: If your PR is very clearly vibe coded slop I am just gonna close it without warning.

Some notes:

• Your PR will be run against the checks here: .github/workflows/ci.yml
• I may at times use AI to review your PRs. A human will still read your PR and have final say

Install

-- lazy.nvim
{
  "cachebag/jumpy",
  config = function()
    require("jumpy").setup({
      provider = "anthropic", -- or "openai", "openrouter"
    })
  end,
}

set your API key: export ANTHROPIC_API_KEY="sk-ant-..." (or OPENAI_API_KEY, JUMPY_API_KEY for openrouter).

Use

Keybind	Action
<leader>j	Open prompt, type your change, hit <CR>
]h / [h	Next / previous hunk
<leader>a	Accept hunk
<leader>x	Reject hunk
<leader>A	Accept all hunks
<leader>X	Reject all hunks
<leader>r	Reprompt the hunk under cursor

License

MIT