Read the diff.
Send the note.
Agent keeps going.

letmediff opens a shareable URL for a curated review of your agent changes. You read the change, scrub the files, leave top-level or line-specific feedback, and the coding session keeps moving while you put the phone down. Diffs are ephemeral: stored for 15 minutes max.

Add the MCP Open a sample review

letmediff.it/diff/a7f3-91c2 3 checkpoints

feedback ⌘ ⏎

▍

cp.04 Adopt valibot schemas 3 files · 2m ago

cp.05 Wire diff socket to MCP 5 files · 1m ago

cp.06 Persist checkpoint cursors 2 files · 12s ago

src/lib/diff/stream.ts +4 −2

@@ src/lib/diff/stream.ts @@

41 41 export function open_stream(session_id: string) {

42 − const es = new EventSource(`/events/${session_id}`);

42 + const es = new EventSource(`/events/${session_id}`, {

43 + withCredentials: true

44 + });

43 45 es.addEventListener('checkpoint', (e) => {

44 − checkpoints.update((c) => [...c, JSON.parse(e.data)]);

46 + const cp = parse(checkpoint_schema, JSON.parse(e.data));

L46 Line note: make parse failures visible to the agent instead of dropping the checkpoint.

47 + checkpoints.update((c) => [...c, cp]);

45 48 });

46 49 return es;

47 50 }

how the handoff works

Four moves.

01
The agent pauses creates checkpoints after every task. Then send you a link.

The letmediff MCP snapshots the working tree, posts the diff, and the session waits. No polling, no copy-paste, no chat-thread archaeology.
mcp.json
```
{
  "letmediff": {
    "command": "npx",
    "args": ["-y", "letmediff"]
  }
}
```
02
A link arrives. You open it.

A review experience specifically built for agent checkpoints. It's not just git. The agent decides how to best show you their changes.

checkpoint rail
cp.01 cp.02 cp.03 cp.04
03
You read. You give feedback.

The sticky feedback sheet stays in reach as you scroll the diff. Leave a session note, or tap a line and attach the comment exactly where it belongs. letmediff carries checkpoint, file, and line context back to the agent.
04
The agent resumes with your note in hand.

Feedback returns to the same MCP session over a kept-open channel. Rinse and repeat until your feature is complete.

why a separate surface

Reviewing the code is the real bottleneck. letmediff tries to build a review experience where agent and human actually cooperate. It's also specifically optimized for mobile so that you can review your code even while working from your phone.

get going

Ask for the work. Review the checkpoints.

Add the MCP server with mcp-add, then keep using your coding agent normally. The agent can create several checkpoints, publish one review URL that contains all of them, then call a wait tool so it stops until you finish reviewing in letmediff.

npx mcp-add   --name letmediff   --type stdio   --command "npx -y letmediff"   --scope project

agent chat letmediff connected

Add line-level comments to the review page and make sure mobile still feels good.

I’ll implement it, create checkpoints as I go, then ask you to review the full diff set.

tool_call letmediff checkpoint Updated landing copy and package name. The package shown in the setup is the package users actually install.

tool_call letmediff checkpoint Added line-level feedback to the diff demo. Line notes are attached in the review, not typed into the final comment.

tool_call letmediff review url 3 checkpoints ready https://letmediff.it/diff/a7f3-91c2

tool_call letmediff wait_for_review Waiting for review in letmediff... The feedback is submitted from the review page, not typed into chat.

Review received. I’ll apply the requested changes and continue from the latest checkpoint.

free to use

Read the diff. Send the note. Agent keeps going.

Four moves.

The agent pauses creates checkpoints after every task. Then send you a link.

A link arrives. You open it.

You read. You give feedback.

The agent resumes with your note in hand.

Ask for the work. Review the checkpoints.

Read the diff.
Send the note.
Agent keeps going.