SparkDown README.md

SparkDown is a small, local markdown editor for the files you and your AI agents write. Run the agent in a split terminal, watch its files change, and review each one in a single queue.

For macOS on Apple Silicon and for Linux. About 4 MB installed on Mac. Free and open source (MIT).

A markdown document opens in Edit view. Command 2 shows it beside its live preview, Command 3 shows only the preview, and Command 4 shows the reading diff with a formatting-only hunk folded. Back in Split, Control backtick opens the terminal drawer, Command D splits it right and Command Shift D splits it down, so three terminals run git status, the diff stat and the log.

The SparkDown window: the file tree with modified files marked M, a markdown document beside its live preview, and three terminal panes below showing git status, the diff stat and the log.
One document in Edit, Split, Preview and Diff (⌘1–⌘4), then the terminal drawer (⌃`) split right and down.

The review loop

Open the folder your agent works in. SparkDown follows the files on disk and keeps a list of what changed, so you read each change once and know what is left.

An agent in the terminal edits docs/search.md, src/indexer.ts and package-lock.json and writes notes/cache-decision.md. Each file flashes in the tree and gets an M or U badge, and the Changes badge counts the files worth reading. The Changes view shows 0 of 3 reviewed, the three files with their line counts, and package-lock.json folded under Probably noise. Space marks search.md reviewed and the down arrow moves on; after cache-decision.md is marked too, the count reads 2 of 3. Then the agent edits search.md again: it goes back to pending with a dot, and the count drops to 1 of 3.

  1. Run the agent in a terminal. Claude Code, Codex, opencode or any other CLI runs in the terminal drawer. Split it right or down, up to four panes. With tmux, sessions survive a restart.
  2. See what it changed. The Changes view lists every changed file with its line counts. Files worth reading come first; lockfiles and generated files fold into a "probably noise" group.
  3. Mark it reviewed. Move with ↑ ↓ and press Space to mark a file reviewed. If the agent edits it again, it goes back to pending with a dot.
The Changes queue: 1 of 4 files reviewed with a progress meter, one file marked as changed again since it was reviewed, a checked file, line counts per file and a collapsed Probably noise group, next to the diff with a Mark reviewed button.
The real Changes view: one reviewed, one changed again since you looked, two still to read.

Read the change, not the churn

The reading diff folds hunks that only reorder imports or reformat code, so what is left is the change itself. A moved block of code is never folded, because order can matter. Switch to Full to see everything.

Reading mode: one import-only hunk and one formatting-only hunk folded, leaving only the change that adds tags to the Entry type and to indexFile.
Reading: two hunks folded, one real change left.
Full mode: the same diff with every hunk, including the reordered imports at the top.
Full: every hunk, including the reordered imports.

What else it does

Works with the agents you use

AgentHow it gets SparkDown's tools
Claude Code, Codex, opencodeAdded for each session when you launch it. Nothing is written to their config.
Gemini CLI, Kiro, GrokAsks once, then installs with the agent's own mcp add.
Cursor CLI, AntigravityAsks once, then adds one entry to the agent's MCP config file.
Any other CLIGets a short prompt that points it at a context file.

Your files stay on your disk. There is no account and no telemetry, and SparkDown never calls an AI vendor's API. The only request it makes on its own is a check on GitHub for updates, at most once a day, with no identifier. You can turn it off in Settings.

Download

Free and open source under the MIT license. Every release is on GitHub.

macOS

Apple Silicon, macOS 10.15 or later. A 3 MB disk image.

Not notarized yet: on first launch, right-click the app and choose Open.

Linux

x86_64 .deb (recommended) or AppImage. Tested on Wayland with Hyprland.

An AUR package is coming.

Windows: a preview build is available. See docs/windows.md on GitHub.