deltaprobe
# deltaprobe š
[](https://github.com/Deepthi-04/deltaprobe/actions/workflows/ci.yml)
[](https://www.npmjs.com/package/deltaprobe)
[](https://registry.modelcontextprotocol.io/v0.1/servers?search=deltaprobe)
[](LICENSE)
> Give your AI assistant x-ray vision into pull requests.
**deltaprobe** is a [Model Context Protocol (MCP)](https://modelcontextprotocol.io) server that lets AI assistants like Claude, Cursor and VS Code Copilot probe GitHub pull requests: list them, read their diffs, summarize changes, flag risky patches and post review comments.
Available on [npm](https://www.npmjs.com/package/deltaprobe) and the [official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers?search=deltaprobe) as `io.github.Deepthi-04/deltaprobe`.
## Tools
| Tool | Status | Description |
|---|---|---|
| `list_open_prs` | ā
| List open PRs in a repo, most recently updated first |
| `summarize_pr` | ā
| Overview of a PR: description, size, file types, most-touched areas, review status and CI results |
| `get_pr_diff` | ā
| Fetch changed files and diff hunks, trimmed to whole hunks within a character budget |
| `check_pr_risk` | ā
| Rule-based risk checks: secrets in added lines (masked), sensitive files, large diffs, missing tests, CI/migration/dependency changes, debug code |
| `post_review_comment` | ā
| Post a review with inline comments; every line is checked against the diff first, with a `dryRun` preview |
## How it works
```
You: "Review PR #42 in acme/app"
ā
ā¼
AI assistant āā(MCP over stdio)āāāŗ deltaprobe āāāŗ GitHub API
ā² ā
āāāāā structured PR data āāāāāāāā
```
The assistant decides which deltaprobe tools to call, deltaprobe fetches and trims the data from GitHub, and the assistant writes the review.
## Quick start
### Prerequisites
- Node.js 20+ (22.12+ for development)
- A GitHub token ([create one](https://github.com/settings/personal-access-tokens)). Optional for reading public repos, but it raises the rate limit from 60 to 5,000 requests/hour. `post_review_comment` needs one with **Pull requests: Read and write** permission.
### Use with Claude Code
```bash
# public repos, read-only (no token)
claude mcp add deltaprobe -- npx -y deltaprobe
# with a token (higher rate limit, private repos, posting reviews)
claude mcp add deltaprobe -e GITHUB_TOKEN=<your-token> -- npx -y deltaprobe
```
### Use with Claude Desktop / Cursor
Add to your MCP config:
```json
{
"mcpServers": {
"deltaprobe": {
"command": "npx",
"args": ["-y", "deltaprobe"],
"env": { "GITHUB_TOKEN": "<your-token>" }
}
}
}
```
Then ask: *"What are the open PRs in modelcontextprotocol/typescript-sdk?"*
## Development
```bash
git clone https://github.com/Deepthi-04/deltaprobe.git
cd deltaprobe
npm install
npm test # run unit tests
npm run build # compile TypeScript to dist/
npm run dev # run the server from source
```
### Project structure
```
src/
āāā index.ts # MCP server entry: registers tools, connects over stdio
āāā github.ts # GitHub API client
āāā diff.ts # hunk splitting, budget fitting, commentable lines
āāā files.ts # file type and area classification
āāā risk/rules.ts # risk rules: each takes the PR's files and returns findings
āāā tools/ # one file per tool
test/ # vitest unit tests
```
## Roadmap
- [x] `list_open_prs`
- [x] `get_pr_diff` with smart truncation
- [x] `summarize_pr`
- [x] `check_pr_risk` rules
- [x] `post_review_comment`
- [x] CI on GitHub Actions
- [x] Publish to npm and the MCP Registry
- [ ] List on Smithery
- [ ] GitHub Action that reviews PRs automatically
## License
[MIT](LICENSE)
TDQS
Scored across 5 tools
Each tool targets a distinct stage of PR review: listing (list_open_prs), overview metadata (summarize_pr), code content (get_pr_diff), heuristic risk scanning (check_pr_risk), and writing feedback (post_review_comment). Descriptions explicitly state sequencing and warn against confusing heuristics with actual code reading, leaving no ambiguity.
All five names follow a consistent snake_case verb_noun pattern (list_, get_, summarize_, check_, post_) with a shared 'pr' domain noun. No mixed conventions or vague verbs.
Five tools map cleanly onto a coherent review workflow with no redundancy or filler. Each tool covers a necessary and non-overlapping step from discovery to posting a review.
The surface covers the core review lifecycle: discover PRs, understand them, read diffs, assess risk, and post comments with inline feedback. Minor gaps remain around reading existing review threads/replies and retrieving CI log details, but agents can work around these.