Tachikoma MCP Server

MCP server with 5 tools for AI agent orchestration.

Overview

Implements tools for:

• Topology-aware orchestration
• Verification loops with grounding
• Skill learning from execution traces
• Graph-based self-healing tool routing
• Enhanced RLM with hierarchical indexing

Designed to work with Tachikoma-Agent-Skills agent.

Features

analyze_topology

Recommends orchestration patterns (sequential/parallel/hierarchical/hybrid) for task routing.

execute_with_verification

Executes tasks with iterative verification and external grounding.

learn_skill_outcome

Records skill execution results for continuous improvement.

query_graph_memory

Queries graph-based memory for intelligent tool selection.

enhanced_rlm_process

Processes large contexts with semantic/adaptive/fixed chunking.

Research Documents

• AdaptOrch: Topology-aware orchestration
• Mirror Loop: Verification loops with external grounding
• SkillOrchestra: Learning from execution traces
• Graph-based Self-healing: Dynamic tool routing
• LycheeCluster: Hierarchical indexing for RLM
• Additional research
• Additional research
• Additional research

Architecture

Modular Python implementation:

tachikoma-mcp-python/
├── src/tachikoma_mcp/
│   ├── server.py
│   ├── models.py
│   └── tools/
├── tests/
├── pyproject.toml
└── run_server.py

Installation

Requires Python 3.14+ and uv:

git clone <repo-url>
cd tachikoma-mcp-python
uv sync

Usage

Run server:

python run_server.py
# or
python -m tachikoma_mcp.server

MCP Registration

Add to agent's MCP config:

{
  "mcpServers": [
    {
      "command": "cd /path/to/tachikoma-mcp-python && python run_server.py",
      "name": "tachikoma-mcp",
      "description": "MCP server for agent orchestration"
    }
  ]
}

Development

Run tests: pytest tests/ -v

Code style: Direct, minimal comments, single-word variables when clear.

License

MIT