/create

name: create
description: Extend the system by creating new MCP servers, agents, or skills. Usage: /create a DeFi protocol tracker or /create an agent for macro analysis
user-invocable: true

Create New Component

Extend the crypto trading desk with a new component based on the user's natural language description: $ARGUMENTS

Workflow

Step 1: Parse Request

Determine what the user wants to create:

• **MCP server** — if the request involves new data sources, APIs, or tools (e.g., "on-chain analytics", "DeFi tracker", "social media monitor")
• **Agent** — if the request involves a new specialist role (e.g., "macro analyst", "DeFi strategist", "on-chain detective")
• **Skill** — if the request involves a new workflow or command (e.g., "multi-coin comparison", "rebalance portfolio", "alert system")

If unclear, ask the user what type of component they want.

Step 2: Research

Delegate to `system-builder` agent: "Research what's needed to create: $ARGUMENTS. Use WebSearch to find relevant public APIs (prefer no-API-key-required). Use WebFetch to read API documentation. Read existing components in the project to understand patterns — read at least 2 files from the relevant directory (mcp-servers/, agents/, or skills/). Read mcp-servers/validators.py for reusable validation. Write a research summary to data/create/{name}-research.md with: APIs found, rate limits, data available, recommended approach."

Step 3: Generate

After research completes, delegate to `system-builder` agent: "Based on the research in data/create/{name}-research.md, generate a new {type} for: $ARGUMENTS. Follow the exact patterns from existing files. Write the component to the correct location:

• MCP server → mcp-servers/{name}.py
• Agent → agents/{name}.md
• Skill → skills/{name}/SKILL.md

Write a creation summary to data/create/{name}-summary.md."

Step 3b: Generate Tests (MCP servers only)

If the component is an MCP server, delegate to `system-builder` agent: "Generate a test file for the new MCP server `mcp-servers/{name}.py`. Read `tests/helpers.py` to understand the `call_tool()` helper. Read at least 2 existing test files (e.g., `tests/test_crypto_data.py`, `tests/test_crypto_exchange.py`) to understand the testing pattern:

• Import with `sys.path.insert(0, str(Path(__file__).resolve().parent.parent / 'mcp-servers'))`
• Use `from helpers import call_tool` wrapper for FastMCP tools
• Mock ALL external API calls (HTTP, CCXT, etc.) — tests must run offline
• One test class per tool, with at least: `test_success` and `test_error_handling`
• Assert `result['status'] == 'success'` on happy path
• Assert `result['status'] == 'error'` on failure path

Write the test to `tests/test_{name}.py`."

Step 4: Integration Guidance

After the component is generated, present:

1. **What was created** — show the file path and a brief description 2. **Review** — show the key parts of the generated file for the user to review 3. **Next steps** — tell the user what they need to do:

For **MCP servers**:

• Add the server configuration to `mcp-servers.plugin.json`
• Add the server name to relevant agents' `mcpServers` list in their frontmatter
• Add the server to the MCP table in `CLAUDE.md`
• Run `uv run --frozen pytest tests/test_{name}.py -v` to verify the tests pass
• Run `/setup` to verify the new server starts correctly

For **agents**:

• Add routing rules in `CLAUDE.md` (which queries should route to this agent)
• Add the agent to the agents table in `CLAUDE.md`

For **skills**:

• Auto-discovered by Claude Code — no configuration needed
• Add to the skills table in `CLAUDE.md` for documentation

Output

Present: 1. Research findings (APIs, feasibility) 2. Generated component (file path + key capabilities) 3. Test results (MCP servers: show pytest output) 4. Integration checklist (numbered steps)