Skip to content

mcp.md

Latest commit

 

History

History
133 lines (97 loc) · 4.63 KB

mcp.md

File metadata and controls

133 lines (97 loc) · 4.63 KB
Raw

Pulsar MCP Server

Pulsar exposes its knowledge base as an MCP (Model Context Protocol) server, making it queryable by Claude, Cursor, or any MCP-compatible client — without custom integration.

Quick Start

Install dependency:

pip install mcp   # requires Python 3.11+

Run (substitute your clone path; the script auto-detects sibling helpers):

python3.11 ~/clawd/scripts/mcp_server.py

Claude Desktop (~/.config/claude/claude_desktop_config.json) — use an absolute path so Claude Desktop can resolve it under your home:

{
  "mcpServers": {
    "pulsar": {
      "command": "python3.11",
      "args": ["/Users/you/clawd/scripts/mcp_server.py"]
    }
  }
}

Override paths:

PULSAR_MEMORY_DIR=/your/memory \
PULSAR_SCRIPTS_DIR=/your/scripts \
python3.11 mcp_server.py

Tools

Signal Tools

Tool Args Description
get_vla_signals days=7, min_rating="🔧" VLA paper signals filtered by recency and rating (⚡ > 🔧 > 📖 > ❌)
get_ai_signals days=7 AI App / Agent daily picks
search_signals keyword, days=30 Full-text search across VLA papers + AI picks

Knowledge Tools

Tool Args Description
get_assumptions domain="all" Active research hypotheses (vla, ai_app, or all)
get_vla_sota days=30 SOTA benchmark records (pass days=0 for all-time)
get_vla_releases days=30 Model/library release events
get_social_intel domain="vla", days=14 Community signals: VLA (structured JSON) or AI (markdown reports)

Meta Tools

Tool Args Description
get_predictions domain="all" Latest biweekly predictions + previous-round results
get_pipeline_health — Last watchdog run status + 7-day signal volume stats

Domain Registry

Tool Args Description
list_domains — All configured Pulsar domains (memory/domains.json) with keys, names, and descriptions
get_domain_config domain Active config (keywords, hypotheses, RSS sources, research directions) for one domain

Search

Tool Args Description
search_memory query, days=60, top=5, source_type="" Semantic search over the 60-day memory window via DashScope text-embedding-v3 + cosine similarity. source_type filters to e.g. social, daily-pick, theory.

Tool count: 12. All tools are read-only; the server never writes to memory files.

Rating Scale

Rating Meaning
⚡ Breakthrough — top-tier paper, major release, or paradigm shift
🔧 Solid — meaningful technical advance, worth tracking
📖 Reference — informational, low immediate impact
❌ Noise — irrelevant or low quality

get_vla_signals defaults to min_rating="🔧", returning ⚡ and 🔧 items only.

Example Queries (via Claude)

What are the top VLA breakthroughs in the last 7 days?
→ get_vla_signals(days=7, min_rating="⚡")

Search for papers about diffusion models in robotics
→ search_signals(keyword="diffusion", days=30)

What hypotheses does Pulsar currently track?
→ get_assumptions(domain="vla")

Is the pipeline healthy? Any failed checks today?
→ get_pipeline_health()

What were last month's predictions and did they come true?
→ get_predictions(domain="all")

What domains is this Pulsar instance tracking?
→ list_domains()

What keywords does the AI app domain use for rating?
→ get_domain_config(domain="ai_app")

What evidence contradicted assumption V-003 last month?
→ search_memory(query="V-003 contradicted", days=30, top=5)

Implementation Notes

  • Runtime: Python 3.11 + mcp 1.26.0 (FastMCP)
  • Transport: stdio (local process, zero network exposure)
  • Memory directory: defaults to ~/clawd/memory/ (the reference deployment path) and is overridable via PULSAR_MEMORY_DIR
  • Scripts directory: auto-detected from mcp_server.py path (overridable via PULSAR_SCRIPTS_DIR); needs to contain _domain_loader.py and semantic-search.py
  • Read-only: the server never writes to memory files
  • AI social intel is stored as dated .md files (_ai_social_YYYY-MM-DD.md); get_social_intel(domain="ai") globs and returns the relevant range
  • search_memory dependencies: semantic index files (memory/semantic-index/{chunks.jsonl,vectors.bin}) must be built by semantic-index-builder.py first; if missing, the tool returns a friendly "index not built" message