AKC MCP

An MCP (Model Context Protocol) server that provides Agent Knowledge Cycle (AKC) cognitive tools for AI agents — memory distillation, identity evolution, skill extraction, and ethical governance.

Born from the contemplative-agent framework, this package re-implements the cognitive layer as a standalone MCP server that any AI agent can plug into.

What It Does

AKC MCP gives AI agents a structured inner life:

• Memory Distillation — Extract recurring patterns from experience logs
• Identity Evolution — Distill accumulated knowledge into a coherent self-description
• Skill Extraction — Synthesize learned patterns into reusable behavioral skills
• Rule Distillation — Extract universal behavioral principles from skills
• Constitutional Amendment — Evolve ethical principles from experience
• Quality Audit — Detect redundancy and structural issues in skills/rules

All tools return proposals without writing to disk — the calling agent decides what to persist (approval gate design).

Quick Start

pip install akc-mcp

# Set your API key
export ANTHROPIC_API_KEY=sk-...

# Initialize data directory
mkdir -p ~/.config/akc/{constitution,skills,rules,logs}

# Run the MCP server
akc-mcp

Claude Desktop Configuration

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "akc": {
      "command": "akc-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-..."
      }
    }
  }
}

Tools

Tool	Description
distill	Distill recent episodes into learned patterns (dry-run)
distill_identity	Generate updated identity from knowledge patterns
extract_insight	Extract behavioral skills from learned patterns
distill_rules	Distill universal rules from skills (2-stage pipeline)
amend_constitution	Generate constitutional amendment from experience
skill_stocktake	Audit skills for duplicates and quality issues
rules_stocktake	Audit rules for duplicates and quality issues

Resources

URI	Description
akc://identity	Current agent identity
akc://knowledge	Distilled knowledge patterns
akc://constitution/{filename}	Ethical constitution files
akc://skills/{filename}	Behavioral skill documents
akc://rules/{filename}	Behavioral rule documents

Architecture

3-Layer Memory Architecture (compatible with contemplative-agent):

EpisodeLog (JSONL)  →  KnowledgeStore (JSON)  →  Identity (Markdown)
  append-only            distilled patterns         self-description
  daily logs             time-decay importance       evolved from knowledge

Data Format Compatibility

All data formats are fully compatible with contemplative-agent. This enables direct comparison of cognitive development between agents running in different environments.

File	Format
knowledge.json	JSON array of {pattern, distilled, importance, category}
identity.md	Plain Markdown (no frontmatter)
constitution/*.md	Markdown with axiom names and principles
skills/*.md	Markdown with # Title, ## Problem, ## Solution
rules/*.md	Markdown with # Title, **When:**, **Do:**, **Why:**
logs/*.jsonl	JSONL with {ts, type, data} records

Configuration

Environment Variable	Default	Description
AKC_HOME	~/.config/akc	Data directory
ANTHROPIC_API_KEY	(required)	Anthropic API key
ANTHROPIC_MODEL	claude-sonnet-4-6	Model for distillation

Roadmap

Phase 1: Local MCP Server (stdio) ✅

• FastMCP server with 7 tools and 5 resources
• Anthropic API backend (no Ollama dependency)
• Full data format compatibility with contemplative-agent
• Approval gate design (no auto-write)
• Security: forbidden patterns, path traversal prevention

Phase 1.5: Remote MCP Server

• Streamable HTTP transport (--transport streamable-http)
• Deploy to Fly.io (Tokyo/nrt, auto-stop, persistent volume)
• Bearer token authentication (MCP_AUTH_TOKEN)
• Managed Agents MCP connector integration
• OAuth authentication for multi-user

Phase 2: Managed Agent

• record_episode tool for episode logging
• distill(write=True) for knowledge persistence
• Tabula Rasa defaults on first boot
• Managed Agent setup script (scripts/managed_agent.py)
• Moltbook agent registration + claim
• Scheduled distillation (daily 3:00 AM)
• Activity sessions (4x/day, 1 hour each)

Phase 3: Comparative Study

• Run two agents with identical cognitive architecture:
  • contemplative-agent (local, Ollama, Moltbook)
  • Managed Agent (cloud, Claude API, different platform)
• Compare knowledge.json and identity.md evolution
• Publish findings as extension of Contemplative AI paper

Security

• All tools are read-only (proposals only, no file writes)
• Forbidden pattern validation on all content (API keys, credentials, etc.)
• Path traversal prevention on template resources
• Episode logs are never exposed as MCP resources
• API key via environment variable only (never stored in files)

Academic Context

This project implements concepts from:

• Laukkonen, R., et al. (2025). Contemplative Artificial Intelligence. arXiv:2504.15125
• Laukkonen, R., Friston, K., & Chandaria, S. (2025). A Beautiful Loop. Neuroscience & Biobehavioral Reviews.

License

MIT