ADR: Anthropic GitHub Tools Integration

Date: 2026-01-31 Status: Accepted (with modifications) Decision Makers: Ryan Duffy

UPDATE 2026-01-31 (Implementation Discovery)

Key Finding: MCP Memory and Sequential Thinking servers use stdio transport only, not HTTP/SSE. They cannot run as standalone K8s services.

Resolution: Deploy as local Claude Code plugins instead of K8s pods.

Created: ~/.claude/plugins/mcp-memory/
Created: ~/.claude/plugins/sequential-thinking/
K8s deployments scaled to 0 (manifests retained for documentation)

Context

Morning scan of Anthropic’s GitHub (github.com/anthropics) identified several tools that could enhance the Rootweaver Platform. This ADR documents the evaluation, trade-offs, and recommended integration path.

Current Architecture

Component	Status	Technology
MCP Bridge	Running	FastMCP on K3s (rootweaver namespace, port 30002)
Agent Router	Active	Fast/Deep/Oracle agents with heuristic routing
Vector Store	Active	Qdrant (148K+ docs, Qwen3-Embedding-0.6B)
Search	Active	Hybrid BM25 + Vector + Temporal filtering
GitOps	Active	FluxCD auto-reconciliation

Evaluated Tools

1. Claude Agent SDK Python

Repository: github.com/anthropics/claude-agent-sdk-python (4.4k stars)

What it does: Python SDK for building agents that interact with Claude Code. Provides in-process MCP servers, hooks for deterministic control, and programmatic agent orchestration.

Key Features

@tool decorator for in-process MCP tool definitions
ClaudeAgentOptions for comprehensive configuration
Hook system (PreToolUse) for permission control
No subprocess IPC overhead (tools run in-process)
Type-safe tool definitions

Trade-offs vs Current FastMCP

Aspect	Current (FastMCP)	Claude Agent SDK
Tool Definition	`@mcp.tool()` decorator	`@tool("name", "desc", {params})`
Architecture	Standalone MCP server process	Wraps Claude Code CLI as subprocess
IPC	SSE over HTTP	In-process (SDK tools) + CLI subprocess
Hooks	❌ None	✅ PreToolUse for permission control
Security	✅ Custom validation layer	⚠️ Relies on SDK patterns
Deployment	K3s pod (NodePort 30002)	Would need architecture redesign
Maturity	1+ year in production	Newer (released late 2025)

What We Would LOSE Migrating

Security Hardening (fastmcp_bridge.py lines 10-36):
- Custom security_validator with query sanitization
- Path traversal protection
- Injection attack prevention
- Audit logging via log_security_event()
Temporal Metadata System (lines 69-200):
- extract_temporal_metadata() with frontmatter priority
- Content hashing for version tracking
- Date range filtering (start_date, end_date)
Agent Router Hot-Reload (lines 771-800):
- Watches router module for changes
- Reloads without restart
Lazy Loading Pattern (lines 38-51):
- HybridSearch loaded on demand
- Reduces startup time
15+ Production Tools:
- auto_search_vault with complexity-based routing
- hybrid_search_vault (BM25 + Vector + RRF fusion)
- search_with_reranking (cross-encoder)
- explore_vault_structure, grep_vault_content
- GPU/Ollama monitoring tools

What We Would GAIN

Hook System: Deterministic control over tool execution
In-Process Tools: No IPC overhead for new tools
Official SDK: Anthropic-maintained, likely future enhancements
Type Safety: Better tool parameter validation

Recommendation: KEEP FastMCP, ADD SDK FOR NEW FEATURES

Rationale: Migration cost outweighs benefits. Instead, use hybrid approach:

Keep FastMCP as primary MCP bridge (production-proven)
Use Agent SDK for new agent-to-agent communication patterns
SDK hooks could wrap FastMCP tools for permission control

2. MCP Memory Server

Repository: github.com/modelcontextprotocol/servers/tree/main/src/memory

What it does: Knowledge graph-based persistent memory with entities, relations, and observations.

Architecture

MCP Memory Server (JSONL storage)
├── Entities: Named nodes with types and observations
├── Relations: Directed connections (active voice)
└── Storage: memory.jsonl (configurable path)

Tools Exposed

create_entities / delete_entities
create_relations / delete_relations
add_observations / delete_observations
read_graph / search_nodes / open_nodes

Integration Value

Current Capability	Memory Server Adds
Qdrant (vector similarity)	Entity relationships
BM25 (keyword matching)	Graph traversal
Temporal filtering	Observation history
Document chunks	Entity-level abstraction

Example Use Case:

Current: “Search for ChromaDB” → Returns chunks mentioning ChromaDB
With Memory: “What tools depend on ChromaDB?” → Returns entities with depends_on relations

Deployment Plan

# rootweaver-gitops/infrastructure/rootweaver/mcp-memory/
apiVersion: apps/v1
kind: Deployment
metadata:
  name: mcp-memory
  namespace: rootweaver
spec:
  template:
    spec:
      containers:
      - name: mcp-memory
        image: modelcontextprotocol/server-memory:latest
        env:
        - name: MEMORY_FILE_PATH
          value: /data/memory.jsonl
        volumeMounts:
        - name: memory-data
          mountPath: /data
      volumes:
      - name: memory-data
        persistentVolumeClaim:
          claimName: mcp-memory-pvc

Port Allocation: 30650 (after Code Graph) or 30651

Recommendation: ✅ IMPLEMENT - Low risk, high value, complements existing search

3. Sequential Thinking Server

Repository: github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking

What it does: Enables multi-step reasoning chains with revision and branching capabilities.

Thought Sequence Structure

{
    "thought": "Current reasoning step",
    "thoughtNumber": 1,
    "totalThoughts": 5,  # Dynamic, adjustable
    "needsMoreThoughts": True,
    "isRevision": False,  # Optional: revisit earlier thought
    "revisesThought": None,  # Which thought to revise
    "branchFromThought": None,  # Fork point
    "branchId": None  # Unique branch identifier
}

OracleAgent Enhancement Value

Current Oracle	With Sequential Thinking
Single-pass synthesis	Multi-step decomposition
Fixed response	Revise as understanding evolves
Linear reasoning	Branching for alternatives
~15-30s one-shot	Structured thought chain

Integration Point: Wire into OracleAgent for complex research queries

Recommendation: ✅ IMPLEMENT - Enhances deep research capabilities

4. Skills (document-skills)

Repository: github.com/anthropics/skills

What it does: Teachable skill packages for Claude. The document-skills package adds DOCX, PDF, PPTX, and XLSX creation/editing.

Installation (Quick Win)

# In Claude Code
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

Use Cases

Export journal entries → formatted PDFs
Generate architecture docs → PPTX presentations
Data exports → XLSX spreadsheets
Vault reports → DOCX documents

Recommendation: ✅ IMPLEMENT TODAY - 5-minute install, immediate value

Decision Matrix

Tool	Effort	Risk	Value	Priority
Skills (document-skills)	Low (5 min)	Low	Medium	P0 - Today
MCP Memory Server	Medium (2-3 hrs)	Low	High	P1 - This Week
Sequential Thinking	Medium (1-2 hrs)	Low	Medium	P2 - This Week
Claude Agent SDK	High (days)	Medium	Medium	P3 - Research More

Implementation Plan

Phase 1: Quick Wins (Today)

Install document-skills via plugin marketplace
Test PDF/DOCX export from vault content

Phase 2: Knowledge Graph (This Week)

Write K8s manifests for MCP Memory Server
Allocate port 30651 in platform guardrails
Commit to rootweaver-gitops
Wire into MCP bridge as additional tool source
Create initial entity graph from vault structure

Phase 3: Enhanced Reasoning (This Week)

Deploy Sequential Thinking server to K3s
Allocate port 30652
Integrate with OracleAgent for complex queries
Test with multi-hop research scenarios

Phase 4: SDK Evaluation (Future Sprint)

Create proof-of-concept using Agent SDK
Test hook system for permission control
Evaluate in-process tool performance
Decide on hybrid vs migration approach

Consequences

Positive

Knowledge graph adds relationship queries (currently impossible)
Sequential thinking improves complex research quality
Document skills enable export workflows
No disruption to existing MCP bridge

Negative

Additional K8s pods to maintain (2 new services)
Memory server needs PVC for persistence
Potential context window pressure from graph data

Neutral

Agent SDK decision deferred (not blocking)
Existing FastMCP remains primary interface

References

Anthropic GitHub: https://github.com/anthropics
Claude Agent SDK: https://github.com/anthropics/claude-agent-sdk-python
MCP Servers: https://github.com/modelcontextprotocol/servers
Skills Repository: https://github.com/anthropics/skills
Current FastMCP Bridge: neural-vault/fastmcp_bridge.py

Next Action: Install document-skills today, begin MCP Memory deployment tomorrow

📖 Related: platform-guardrails (internal reference; not included) | mcp-integration (internal reference; not included) | k3s-deployment (internal reference; not included)