Package Exports

devxira
devxira/build/cli.js

This package does not declare an exports field, so the exports above have been automatically detected and optimized by JSPM instead. If any package subpath is missing, it is recommended to post an issue to the original package (devxira) to support the "exports" field. If that is not possible, create a JSPM override to customize the exports field for this package.

Readme

Devxira

MCP server for developer workflows — PR review, daily standup reports, task management, and WordPress plugin research.

Features

Git — Status, diff, branch, commit, push, log (local stdio mode only)
WordPress — Search and inspect plugins from wordpress.org
Workflow prompts — Built-in workflow templates for daily reports, PR reviews, task implementation, plugin research, and QA testing
Companion server support — Integrates with GitHub, ClickUp, and Playwright MCP servers for extended workflows

Installation

Claude Code

claude mcp add devxira -- npx devxira

Companion Servers (optional)

Install these alongside devxira for full workflow support:

# GitHub — for PR review, code search, and GitHub operations
claude mcp add --transport http github https://api.githubcopilot.com/mcp/ -H "Authorization: Bearer YOUR_GITHUB_PAT"

# ClickUp — for daily reports and task management
claude mcp add clickup --transport http https://mcp.clickup.com/mcp

# Playwright — for browser-based QA testing
claude mcp add playwright -- npx @playwright/mcp@latest

Claude Desktop

Add to your config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "devxira": {
      "command": "npx",
      "args": ["devxira"]
    }
  }
}

VS Code / Cursor

Add to .vscode/mcp.json in your workspace:

{
  "servers": {
    "devxira": {
      "command": "npx",
      "args": ["devxira"]
    }
  }
}

Windsurf

Add to ~/.windsurf/mcp.json:

{
  "mcpServers": {
    "devxira": {
      "command": "npx",
      "args": ["devxira"]
    }
  }
}

OpenAI Codex

codex --mcp-server "npx devxira"

Available Tools

Git (6) — stdio mode only

Tool	Description	Read-Only
`git_status`	Show working tree status	Yes
`git_diff`	Show staged/unstaged/ref changes	Yes
`git_create_branch`	Create and checkout a new branch	No
`git_commit`	Stage files and commit	No
`git_push`	Push to remote	No
`git_log`	Show recent commit history	Yes

WordPress (2)

Tool	Description	Read-Only
`wordpress_get_plugin`	Get plugin details from wordpress.org	Yes
`wordpress_search_plugins`	Search plugins by keyword	Yes

Workflow Prompts

Devxira includes workflow prompts that guide Claude through multi-step developer tasks. Prompts that need companion servers will automatically detect if they're missing and show install instructions.

Prompt	Description	Companion Server
`test_connection`	Verify all integrations are working	—
`daily_report`	Generate and send daily standup report	GitHub, ClickUp
`audit_pr`	Single-pass deep audit → severity-rated bug report	GitHub
`pr_feedback_loop`	Incremental Codex + Claude loop: review → verify → fix → re-review. A critic agent pushes back on false positives, a fixer agent applies verified fixes, then the loop repeats until clean or the iteration cap (default 5) fires. Best for small/medium PRs and fast feedback.	GitHub (PR mode only)
`pr_bulk_review_loop`	Bulk-discovery multi-agent loop: parallel specialized discovery passes (no fixing) → consolidate + cluster by root cause → evidence-grounded specialist critics on the FULL list → architect a coordinated fix plan → apply all fixes in dependency order → final verify. Best for medium/large PRs and root-cause / cross-cutting bugs (one fix closes N findings). Honors CLAUDE.md, `.claude/skills/*/SKILL.md`, and persists false-positive learnings across runs.	GitHub (PR mode only); Playwright (auto, for UI changes)
`implement_task`	End-to-end ClickUp task implementation with PR	GitHub, ClickUp
`research_plugin`	WordPress plugin market analysis and MVP planning	—
`build_plugin`	Generate step-by-step AI build plan from a research report	—
`qa_test`	Browser-based QA testing with screenshot evidence	Playwright (+ GitHub if PR provided)

Picking between the two PR-review loops

Aspect	`pr_feedback_loop` (incremental)	`pr_bulk_review_loop` (bulk)
Loop shape	small bite (find few → verify → fix → re-find)	one big sweep (find all → verify all → fix all → final verify)
Best for	small/medium PRs, fast feedback, well-known codebase	medium/large PRs, security audits, root-cause / pattern bugs, unfamiliar codebases
Catches	individual bugs as they surface	cross-cutting patterns (one fix closes 5 findings)
Wall-clock	longer (sequential rounds)	shorter (parallel discovery + fix in batches)
Risk	misses cross-file patterns, can oscillate	longer time-to-first-fix; bugs may interact under batch fix
Args	`pr`, `max_iterations`, `mode`	`pr`, `discovery_passes`, `outer_iterations`, `mode`, `verify_with_browser`, `precision`

Research-backed design (`pr_bulk_review_loop`)

The bulk loop's design is grounded in published code-review-tool research (CodeRabbit, Cursor Bugbot V1→V2 retrospective, CriticGPT, AutoReview FSE 2025, the Multi-Agent Code Verification paper, and the bug-bash methodology). Highlights:

Specialist critics over generic — domain-specific critics (Correctness / Security / Performance / UI / Backend-Tenancy / Type-safety) detect 3× more bugs and produce 87% fewer false positives than generic critics, because their detections are statistically independent.
"Evidence not vibes" (CodeRabbit) — every critic verdict requires executable proof (grep / ast-grep / quoted code snippets / browser screenshots). Pure-opinion verdicts are auto-downgraded to NEEDS_INFO.
Asymmetric thresholds, not voting — Security critic alone marking VERIFIED blocks; Correctness needs two specialists or one + concrete reproduction; UI requires a screenshot or DOM excerpt. This prevents the "80+ agents unanimously endorse a non-existent vulnerability" failure mode (real, documented).
Dynamic per-finding context lookup (Cursor Bugbot V2 lesson) — critics grep and Read the actual code per finding rather than trusting an inlined excerpt. Resolution rate climbed from 52% → 70% when Cursor moved from pre-stuffed context to dynamic.
Persistent learnings (CodeRabbit-style) — refuted findings get appended to .claude/review-learnings.md. Future runs auto-skip patterns matching that file (overridable with new evidence). Cross-PR memory is the single biggest source of false-positive reduction over time.
Architect / Editor split (Aider) — one subagent plans the fix batch (with dependency ordering and cluster-fix detection), separate fixer subagents apply patches per file-group.
Browser verification uses Playwright MCP via the bundled qa_test prompt (or codex exec with Playwright MCP wired). Codex CLI does not have a native browser plugin — @Browser is a Codex App (GUI) feature only.

Note on naming: the previous review_pr prompt was renamed to audit_pr in v1.2.0 to make room for the loop tools and to avoid colliding with project-level /review-pr slash commands. pr_bulk_review_loop was added in v1.3.0.

Authentication

All external authentication is handled by companion MCP servers:

GitHub — Requires a GitHub Personal Access Token (PAT). Create one at github.com/settings/tokens with repo and read:user scopes.
ClickUp — Authenticates via browser on first use through ClickUp's official MCP server.
Playwright — No authentication needed (runs locally).

Development

Prerequisites

Node.js >= 18
npm

Setup

git clone https://github.com/abuldev/devxira.git
cd devxira
npm install

Build

npm run build

Run

# stdio mode (for MCP clients)
npm start

# HTTP mode (for remote deployment)
npm run start:http

Test

npm test

Test with MCP Inspector

npx @modelcontextprotocol/inspector node build/cli.js

License

MIT

devxira