Phabricator MCP Server

Standalone MCP server for Phabricator, kept outside the devel repo so hg stays clean.

Setup

1. Copy env file and fill credentials:

   cp .env.example .env
   # Edit .env with your Phabricator API token

2. Install deps (already done in venv):

   python3 -m venv venv
   ./venv/bin/pip install mcp requests python-dotenv

Related MCP server: Sentinel Core Agent

Run

./run.sh

Or manually:

source venv/bin/activate
python3 server.py

Tools exposed

Phabricator API

• search_tasks – query tasks by status / priority / assigned / text

• get_task – fetch a single task by ID with full description, media, and comments

• get_task_comments – fetch only comments / transactions for a task

• create_task – create a new maniphest task

• edit_task – update an existing task or add a comment

• search_revisions – search differential revisions

• get_revision – fetch a revision by ID

• get_revision_diff – fetch raw diff for a revision

• get_unbreak_tasks – list open unbreak tasks

• get_projects – search Phab projects

• get_project_members – list members of a project

• search_users – search Phab users

• get_file_info – get metadata for a file / image by PHID

• query_phids – resolve arbitrary PHIDs to objects

• ping – health-check connectivity

RAG (Semantic Search)

• phab_search_rag – semantic search over indexed tasks and revisions

• phab_ask – natural-language Q&A with source attribution

RAG Architecture

Phabricator API                    ChromaDB Vector Store
     │                                      ▲
     ▼                                      │
┌─────────────┐    ┌─────────────┐    ┌──────────┐
│  Extractor  │───▶│   Chunker   │───▶│ Indexer  │
│  (Conduit)  │    │ (tiktoken)  │    │(metadata)│
└─────────────┘    └─────────────┘    └──────────┘
                                            ▲
┌─────────────┐    ┌─────────────┐          │
│ Diff Parser │───▶│  Embedder   │──────────┘
│(files/hunks)│    │(OpenAI 3-sm)│
└─────────────┘    └─────────────┘

Query Flow:
┌──────────┐    ┌──────────┐    ┌─────────┐    ┌──────────┐
│  User    │───▶│ Embedder │───▶│  Search │───▶│  LLM     │
│  Query   │    │ (OpenAI) │    │(ChromaDB)│    │(Anthropic│
└──────────┘    └──────────┘    └─────────┘    │ Claude)  │
                                                └──────────┘

Tech Stack

• Embeddings: OpenAI text-embedding-3-small (1536 dims)

• Vector DB: ChromaDB with cosine similarity + metadata filtering

• LLM: Anthropic Claude Haiku for RAG synthesis

• Chunking: tiktoken-based with 512-token chunks, 50-token overlap

• Diff Metadata: Files changed, hunk context lines (function/class names)

RAG Setup

1. Add API keys to .env:

   OPENAI_API_KEY=sk-...          # Required for embeddings
   ANTHROPIC_API_KEY=sk-ant-...   # Required for RAG Q&A

2. Run the full index (one-time):

   source venv/bin/activate
   python scripts/full_index.py

   • ~47K tasks + ~38K revisions = ~116M tokens

   • Cost: ~$2.32 (OpenAI embeddings)

   • Time: ~16 hours (fast mode) or ~30 hours (full mode)

3. Fast mode (recommended for first run):

   python scripts/full_index.py --skip-comments --skip-diffs

   • Indexes titles, descriptions, summaries only

   • ~4–6x faster than full mode

   • Resume capability: re-run without flags later to enrich with comments/diffs

4. Test a subset first:

   python scripts/full_index.py --task-limit 10 --rev-limit 10

Incremental Sync

After the initial index, run incremental sync to pick up new and modified tasks/revisions:

python scripts/incremental_sync.py

This fetches only objects modified since the last sync (stored in last_sync.json) and updates the vector store. Use --dry-run to preview what would be synced without making changes.

Cron setup (daily sync)

# Add to crontab (crontab -e)
# Replace /path/to/mcp-phab with your project directory
0 2 * * * cd /path/to/mcp-phab && venv/bin/python scripts/incremental_sync.py >> /tmp/phab_sync.log 2>&1

RAG Usage

Via MCP Tools

phab_search_rag(query="find bugs about email signups", limit=5)
phab_ask(query="What caused the AttributeError in matching filters?")

Via CLI (direct Python)

python -c "
from rag.embedder import Embedder
from rag.indexer import Indexer
emb = Embedder()
idx = Indexer()
results = idx.search(emb.embed(['email signup bugs'])[0], limit=3)
for r in results:
    print(r['metadata']['source_uri'], r['metadata']['title'])
"

Claude Desktop config

Add to your Claude Desktop MCP settings (~/.config/claude/mcp-config.json or similar). Replace /path/to/mcp-phab with your actual project directory and fill in your credentials:

{
  "mcpServers": {
    "phab": {
      "command": "/path/to/mcp-phab/venv/bin/python3",
      "args": ["/path/to/mcp-phab/server.py"],
      "env": {
        "PHABRICATOR_URL": "https://phabricator.example.com",
        "PHABRICATOR_API_TOKEN": "your-api-token-here",
        "TRANSPORT": "stdio"
      }
    }
  }
}