redline.dev / v0.1 / open source
A document-based IDE
Code IDEs center your source files. Redline centers the document — the plan, the spec, the research brief — and builds the IDE around it: a track-changes editor, an embedded browser, terminals, and a Claude Code agent on every surface. It’s where the thinking happens before the code.
Free & open source · Apache-2.0 · macOS (Apple Silicon + Intel) · Local-first, no account, no telemetry.
Fig. 01 / The document at the center
The IDE around the document
surfaces
Everything you do before the code — research, drafting, planning, review — in one local window, each surface paired with its own Claude Code agent.
An embedded browser where every tab has its own Claude agent that reads and drives the page — open links, click, extract. Pin what matters, and a Mission agent weaves your tabs and pins into a brief.
browser · page agents · missions
The Prompt Drafter turns loose research into a sharp prompt or spec — the seed of the plan, written in a real editor instead of a chat box.
rich-text · brief → prompt
Claude Code's plan opens in a track-changes editor instead of a terminal prompt. Mark it up the way a lawyer redlines a contract — inline, in place.
tiptap · tracked marks
Comments, questions, and structural edits anchored to the exact line. Claude posts its suggestions right in the document — Accept or Reject in your hands — and every round is resolved.
ask/revise · accept/reject
How it works
protocol
The plan-review loop is the spine — here’s the round-trip between Claude Code and the document.
- 01Intercept
Intercept — Redline installs a PreToolUse hook on ExitPlanMode. When Claude Code finishes planning, the plan is POSTed to a local daemon and the session is held open while you review.
- 02Redline
Redline — The plan opens in a track-changes editor. Type edits directly into the text, and attach edits, feedback, questions, and structural changes exactly where they apply.
- 03Submit
Submit — Two modes: Ask sends questions only and the plan body comes back unchanged. Revise sends your full markup and drives a new revision.
- 04Resolve
Resolve — The next revision arrives with a resolution attached to every comment. Accept each one, or reopen it for the next round. Approve when it's right.
Modes
interception
Every plan blocks and waits for your review. Nothing proceeds until you approve or send revisions.
block & review
A decision window with a countdown. Claim it to review in full, or let it auto-approve and stay in flow.
countdown · auto-approve
Kill switch. Everything passes straight through, nothing captured.
full passthrough
Switch from the header or the tray icon. Redline sits in your tray and shows pending reviews at a glance.
A full workspace
capabilities
Real PTY-backed shells and a sidebar file tree with a fast read-only viewer — run Claude Code inside Redline and review the plan next to the code it touches.
xterm · pty · file tree
Open a read-only side conversation with Claude on any comment. Ask follow-ups about the plan without disturbing the held session.
fork agent · read-only
Browse every version of a plan, restore an earlier one, and export any revision as clean Markdown or DOCX.
v1…vN · md · docx
Talk through a plan hands-free — a voice agent that holds the session, push-to-talk dictation, and read-aloud TTS (on-device or cloud).
voice · dictation · tts
Review state persists through crashes and restarts — the live document in Yjs + IndexedDB, sessions and comments in SQLite.
yjs · sqlite
Redline installs its own Claude Code hook and skill on first run. Both are plain files you can inspect, and the hook can be paused from inside the app.
hook · skill
Why Redline
rationale
Redline was built by a lawyer who wanted to mark up plans the way lawyers redline contracts: tracked changes, margin comments, and nothing accepted until it’s resolved. The underlying bet is that the cheapest place to catch a mistake is upstream of the code — in the research, the draft, and the plan. So Redline gives that work a real IDE.
- vs. approving in the terminal
- In the terminal, your only response to a plan is one undifferentiated prose message against a wall of text. In Redline you comment on §B.p2, not on scrollback.
- vs. just reading the plan
- You edit it directly. Insertions and deletions appear as tracked marks and return to the agent verbatim.
- vs. accept/undo tools
- Nothing happens until you approve. Every comment gets an explicit resolution in the next revision, and nothing is accepted until you accept it.
- Local-first
- No account, no telemetry. The daemon binds to 127.0.0.1 and your documents live on your machine. Redline drives your own local Claude Code — only the surfaces that inherently reach out (the web browser, optional cloud voice) touch the network.
Install
build from source
No prebuilt binaries yet — Redline is an early release (v0.1) and builds from source. macOS only for now; Windows and Linux are on the roadmap.
prerequisites
- ▸ macOS 11 or later (Apple Silicon or Intel)
- ▸ Claude Code
- ▸ Node.js ≥ 20 and npm
- ▸ Rust (stable, via rustup)
- ▸ Xcode Command Line Tools
npm run redline builds the app, installs it into /Applications, and launches it — from then on, open it like any Mac app. Update with git pull && npm run redline. On first launch, Redline offers a one-click install of its Claude Code hook and plan-revision skill.
git clone https://github.com/sersiousSenpai/redline.git cd redline npm install npm run redline
/repo
Redline is open source.
Read the code, file issues, send patches. No telemetry to audit because there isn’t any.
Tauri 2: a React 19 frontend and a Rust backend embedding an axum daemon, a SQLite session store, PTY shells, and native browser webviews. Agents run as your own local Claude Code, driven over a 127.0.0.1 bridge. Apache-2.0 licensed — your first pull request triggers a quick CLA.
The cheapest place to catch a mistake is the diff the plan.
Free & open source · Apache-2.0 · macOS · Local-first.