Differ

A local code review tool for AI-generated code.

Why?

Reviewing agent-generated code changes has friction:

• Jumping between files to find what changed
• No persistent place to leave feedback
• Copy-pasting feedback back into agent conversations
• Multiple agents may work on the same codebase simultaneously

Differ provides a GitHub PR-like review UX, but local and integrated with AI agents via MCP. You review changes in a web UI, leave inline comments, and agents can see your feedback and respond—all in real-time.

How It Works

1. Agent starts work, calls get_or_create_session to join/create a review session
2. Agent creates/modifies files. Any git tracked files will be automatically shown, and the agent can call register_files to add new files to the review
3. You open the web UI, see the diff, leave inline comments
4. Agent calls get_pending_feedback to see your comments
5. Agent addresses feedback, replies to comments, marks them resolved
6. Repeat until you're satisfied, then push

Sessions are tied to a (project, branch) pair. Multiple agents on the same project+branch share a session, so comments are visible to all participants.

Features

• GitHub-style diff viewer - Split or unified view modes with syntax highlighting
• MCP integration - AI agents can register files, read comments, and respond
• Real-time updates - SSE-based live sync between UI and agents
• Session management - Track multiple review sessions across projects

Quick Start

# Start dev environment (compiles + watches + runs server)
npm install
clj -M:dev

Or using npm/npx:

npm install
npx shadow-cljs watch server ui

Open http://localhost:8081 for development (with hot reload) or http://localhost:8576 for production.

Production Build

# Using clj
clj -M:build

# Or using npx
npx shadow-cljs release server ui

# Start the server
node target/server.js

MCP Integration

Differ exposes an MCP endpoint at http://localhost:8576/mcp for AI agent integration.

Note: The MCP server implements the OAuth flow for clients that require it, but accepts all requests without actually verifying credentials. This is intentional as local services are assumed to be secure. Don't expose it to the internet.

Claude Code

Add to your MCP settings file (~/.claude/settings.json or project-level .claude/settings.json):

{
  "mcpServers": {
    "differ-review": {
      "type": "http",
      "url": "http://localhost:8576/mcp"
    }
  }
}

See Claude Code MCP docs for more details.

Other Agents

The endpoint accepts standard MCP JSON-RPC over HTTP POST. Any agent that supports MCP can connect by pointing to the URL.

Available MCP Tools

Tool	Description
get_or_create_session	Get or create a review session for a repo
list_sessions	List all active review sessions
register_files	Add files to the review set
unregister_files	Remove files from the review set
get_review_state	Get current session state with all comments
get_pending_feedback	Get unresolved comments
add_comment	Add a comment or reply
resolve_comment	Mark a comment as resolved
unresolve_comment	Reopen a resolved comment
submit_review	Finish your review with an optional summary comment
get_session_diff	Get the diff content for a session
get_file_versions	Get original and modified versions of a file
get_context	Get session context (path, branch, PR info)
list_directory	List directory contents at a ref
get_file_content	Get file content at a ref
get_history	Get commit/change history for the session
create_pull_request	Push branch and create a GitHub PR

Slash Commands

Differ includes Claude Code slash commands for reviewing and fixing code:

• /review - Review current uncommitted changes, add comments via differ
• /fix-review - Fix all unresolved review comments

These are available automatically when working in the differ project.

Using globally (all projects)

To use these commands in any project, copy them to your global commands directory:

cp .claude/commands/*.md ~/.claude/commands/

Then /review and /fix-review will work in any repo where the differ MCP server is configured.

Review stop hook

In .claude/hooks/pr_review_loop.py there is a script that can be used to force agents to go through a proper internal review -> GitHub PR -> PR closed flow, by prompting the agent to push all changes to a proper PR, and then wait until it's closed. To use it, add the following to the "hooks" section of the appropriate settings.json file:

    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python3 shared/hooks/pr_review_loop.py",
            "timeout": 43200
          }
        ]
      }
    ]

This will not run if the folder in which Claude Code is running is not a git repo, or if the PR_REVIEW_LOOP_DISABLED env variable is set to 1, true or yes. By default, once all changes have been pushed etc., the script will wait up to 12h polling for unresolved comments on the PR, after which it will just let the agent continue. The hook definition has a timeout (default is 30s) after which it will just continue. So the lower of these two values limits how long the agent will wait for feedback.

If the PR is closed, and there are no local changes, the hook will ask the Claude session to commit sepukku for doing such a good job.

Configuration

Edit resources/config.edn to customize settings:

{:port 8576

 ;; File display thresholds (client)
 :large-file-threshold 50000      ;; Characters - files larger require explicit load
 :line-count-threshold 400        ;; Diff lines - more than this requires explicit expand
 :context-expand-size 15          ;; Lines to expand at a time

 ;; Watcher settings (server)
 :watcher-debounce-ms 300         ;; Debounce file change events

 ;; Push whitelist - controls which repos/branches can be pushed via create_pull_request
 ;; Empty map = all repos/branches allowed
 ;; Example: {"owner/repo" ["feature/*" "fix/*"], "myorg/*" ["*"]}
 :push-whitelist {}}

Push Whitelist

The create_pull_request tool can be restricted via a whitelist in config.edn:

{:push-whitelist {"owner/repo" ["feature/*" "fix/*"]
                  "myorg/*" ["*"]}}

• Empty whitelist (default): All repos/branches allowed
• Repo keys: Can use wildcards like "owner/*" to match any repo from that owner
• Branch patterns: Can use wildcards like "feature/*" to match branches
• "*" matches anything

The whitelist can also be set via the PUSH_WHITELIST environment variable as JSON:

export PUSH_WHITELIST='{"owner/repo": ["main", "develop"]}'

Environment variables:

• PORT - Server port (default 8576)
• DIFFER_URL - Base URL for OAuth callbacks (e.g., http://localhost:8576)