TechKnowmad MCP

A monorepo of production-grade Model Context Protocol servers for structured thinking, code generation, graph analysis, and agent observability. Each server exposes a focused set of tools consumable by Claude and any MCP-compatible AI assistant.

---

Features

• Adversarial idea validation — stress-test concepts with devil's advocate, skeptic, and market-reality modes before committing resources
• Hallucination detection — store ground-truth anchors and validate AI claims against them at call time
• Edge-case benchmarking — auto-generate security, robustness, and boundary-condition test suites from a function spec
• Polyglot code scaffolding — emit typed code skeletons, JSON schemas, and test stubs across multiple languages from a single spec
• Knowledge graph construction — build, query, and compute centrality metrics on in-memory directed graphs
• Agent execution tracing — instrument multi-step agent workflows with timestamped step logs for replay and audit

---

Servers

Server	Tools	Purpose
idea-killer-mcp	kill_idea, stress_test_idea, find_fatal_flaws	Adversarial critique
ground-truth-mcp	store_ground_truth, validate_claim, compare_outputs	Fact validation
edgecraft-benchmark-mcp	generate_edge_cases, run_benchmark, compare_benchmarks	Edge-case generation
forge-generate-mcp	generate_code, generate_schema, generate_tests	Code/schema scaffolding
graph-forge-mcp	create_graph, query_graph, compute_centrality	Knowledge graphs
trace-agent-mcp	start_trace, log_step, get_trace	Agent tracing
phyloid-evolve	evolve_population, compute_phylogeny, mutate_individual	Evolutionary computation
negativa-score	score_negatives, rank_by_downside, compute_risk_profile	Risk scoring

---

Quick Start

Requirements: Python 3.12+, uv or pip

git clone https://github.com/techknowmad/techknowmad-mcp.git
cd techknowmad-mcp
pip install -e ".[dev]"

Run all tests:

pytest -v

Run a single server:

python servers/idea-killer-mcp/server.py

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

{
  "mcpServers": {
    "idea-killer": {
      "command": "python",
      "args": ["/path/to/techknowmad-mcp/servers/idea-killer-mcp/server.py"]
    }
  }
}

---

Architecture

techknowmad-mcp/
├── pyproject.toml              # Root: Python 3.12, pytest, hatchling
└── servers/
    └── {server-name}/
        ├── server.py           # MCP server — tool registration + handlers
        ├── pyproject.toml      # Per-server package config
        ├── SKILL.md            # Tool reference and usage examples
        └── tests/
            └── test_server.py  # Async pytest suite

Each server follows a consistent pattern:

1. Tool registration — @app.tool() decorators declare name, description, and input schema
2. Async handlers — handle_list_tools() and handle_call_tool() implement the MCP protocol
3. JSON responses — all results wrapped in types.TextContent as serialized JSON
4. No external state — in-memory only; suitable for stateless agent sessions

Adding a new server: copy an existing server directory, rename it, update pyproject.toml, implement tools in server.py, add tests, and register in Claude Desktop config.

---

Contributing

See CONTRIBUTING.md for the full workflow (fork → feature branch → tests → pytest -v → ruff check → PR). All PRs must include tests and pass CI.

---

License

MIT

---

Built by TechKnowMad Labs