MCP Refactor Progress#

Branch: mcp-refactor#

Completed ✅#

Phase 1: Foundation#

1. Cloned and studied reference projects

   • sandbox/prefect-mcp-server - Learned PydanticAI + MCP patterns
   • Understood how MCP servers work as toolsets for PydanticAI agents

2. Created simplified memory system (src/bot/memory.py)

   • Single SQLite database (threads.db)
   • Plain text storage - no embeddings, no vector search
   • Two tables:
     • threads - Full conversation history per thread (JSON)
     • user_memories - Simple facts about users
   • Completely interpretable - you can open the db and read everything

3. Integrated ATProto MCP server

   • Copied from .eggs/fastmcp/examples/atproto_mcp → src/bot/atproto_mcp
   • Updated settings to use existing env vars (BLUESKY_HANDLE, etc.)
   • Server provides tools: post(), like(), repost(), follow(), search(), create_thread()

4. Created MCP-enabled agent (src/bot/agent.py)

   • PydanticAI Agent with ATProto MCP tools as a toolset
   • Loads personality from personalities/phi.md
   • Integrates with memory system
   • Returns structured Response (action, text, reason)

5. Updated dependencies

   • ✅ Added: fastmcp>=0.8.0, websockets>=15.0.1
   • ❌ Removed: turbopuffer, openai (no longer needed for memory)

What Changed#

Before (Complex)#

• Memory: TurboPuffer + OpenAI embeddings + semantic search
• Agent: Custom response generator with manual action interpretation
• AT Protocol: Direct client calls scattered throughout codebase
• Personality: Dynamic loading from TurboPuffer
• Self-modification: Complex approval system with DM workflow

After (Simple)#

• Memory: SQLite with plain text (interpretable!)
• Agent: PydanticAI with MCP tools (agent decides actions)
• AT Protocol: MCP server provides all tools
• Personality: Static file loading
• Self-modification: Removed (cruft)

How It Works Now#

# Create agent with memory
memory = Memory()
agent = PhiAgent(memory)

# Process a mention
response = await agent.process_mention(
    mention_text="hey phi!",
    author_handle="user.bsky.social",
    thread_uri="at://did/post/123"
)

# Agent returns: Response(action="reply", text="...", reason="...")
# If action is "reply", agent can call MCP tool: post(text="...", reply_to="...")

The agent has access to all ATProto MCP tools and can decide:

• Should I reply, like, or ignore this?
• If replying, what should I say?
• Should I use other tools (repost, follow, etc.)?

Next Steps#

Phase 2: Integration (Not Started)#

1. Update src/bot/main.py to use new agent
2. Simplify src/bot/services/notification_poller.py
3. Remove old response_generator.py
4. Test end-to-end

Phase 3: Cleanup (Not Started)#

1. Delete cruft:
   • src/bot/ui/ (context visualization)
   • src/bot/personality/editor.py (approval system)
   • src/bot/core/dm_approval.py
   • src/bot/memory/namespace_memory.py
   • src/bot/agents/anthropic_agent.py (replaced by agent.py)
2. Update database.py to remove approval tables
3. Update tests
4. Update README.md and documentation

Phase 4: Verification (Not Started)#

1. Run the bot and test mentions
2. Verify thread memory works
3. Verify user memory works
4. Ensure online/offline status still works

Testing#

Test script created: sandbox/test_new_agent.py

uv run python sandbox/test_new_agent.py

Key Files#

New#

• src/bot/memory.py - Simple SQLite memory
• src/bot/agent.py - MCP-enabled PydanticAI agent
• src/bot/atproto_mcp/ - ATProto MCP server (vendored)

Modified#

• pyproject.toml - Updated dependencies

To Be Deleted#

• src/bot/memory/namespace_memory.py
• src/bot/agents/anthropic_agent.py
• src/bot/response_generator.py
• src/bot/ui/
• src/bot/personality/editor.py
• src/bot/core/dm_approval.py

Philosophy#

Before: Over-engineered for capabilities we might want someday After: Simple, working, interpretable system that does what we need today

The memory is now something you can:

1. Open with any SQLite browser
2. Read and understand immediately
3. Debug by just looking at the tables
4. Migrate or export trivially

No more:

• Vector embeddings you can't see
• Complex namespace hierarchies
• Approval workflows for every personality change
• Multiple overlapping memory systems