@cpsagentkit/mcp-server

Model Context Protocol server exposing Agent Workbench for Copilot Studio's Copilot Studio knowledge, parsing, and assessment tools to any MCP-aware LLM client (Claude Desktop, VS Code, Cursor, custom agents).

Install

# One-shot via npx (no install)
npx @cpsagentkit/mcp-server --transport=stdio

# Or install globally
npm install -g @cpsagentkit/mcp-server
cpsagentkit-mcp --transport=stdio

Transports

• stdio (default): for MCP clients that spawn the server as a subprocess
• HTTP (Streamable): for clients that connect to a long-running server

cpsagentkit-mcp --transport=stdio
cpsagentkit-mcp --transport=http --host=127.0.0.1 --port=3333

The HTTP endpoint is POST /mcp and follows the MCP Streamable HTTP spec.

Claude Desktop / MCP client config

Add to your claude_desktop_config.json (or equivalent):

{
  "mcpServers": {
    "cpsagentkit": {
      "command": "npx",
      "args": ["-y", "@cpsagentkit/mcp-server", "--transport=stdio"]
    }
  }
}

Tools

The server registers the following tools:

Knowledge retrieval

• cps_list_knowledge_topics — lists every bundled CPS knowledge and best-practice document. Optional category filter (knowledge | bestpractices).
• cps_get_knowledge — returns the full markdown body of a knowledge document by slug.
• cps_get_best_practice — returns the full markdown body of a best-practice document by slug.

Workspace parsing

• cps_detect_project_state — snapshot of a workspace (initialised? spec/arch present? agents cloned?).
• cps_list_agents — finds all cloned CPS agent folders in a workspace.
• cps_parse_agent — reads one cloned agent folder into a structured snapshot (settings, topics, actions, knowledge).
• cps_parse_solution — reads an exported CPS solution folder (solution.xml + botcomponents/) into agent snapshots plus metadata.
• cps_find_solution_folders — scans a directory for exported CPS solution folders.

Assessment

• cps_validate_tool_description — lint-checks a tool/topic/agent description against CPS prompt-engineering rules.
• cps_compose_review_prompt — assembles a complete solution review prompt from workspace + bundled rules; scope can be full | prompts | descriptions | architecture.

Build context

• cps_generate_topic_scaffolds — extracts topic scaffolds from architecture.md.
• cps_detect_dataverse_mcp — checks for Dataverse MCP configuration in .vscode/mcp.json.

Resources

Every bundled knowledge and best-practice document is also registered as an MCP resource with URI cpsagentkit://<category>/<slug>, so clients with resource UIs can browse them without invoking tools.

License

MIT — see LICENSE.

Hosted endpoint

A public anonymous instance is hosted on Azure Container Apps:

https://ca-mcp-6frsoq6bw5vuk.salmonstone-8b839f60.uksouth.azurecontainerapps.io/mcp

(Once DNS is configured, https://cpsagentkit.codeworks.app/mcp will resolve to the same endpoint.)

The hosted instance runs with MCP_HOSTED=1, which exposes only the tools that don't need access to your local workspace: cps_list_knowledge_topics, cps_get_knowledge, cps_get_best_practice, cps_validate_tool_description, cps_parse_prompt_config, cps_build_prompt_update. For workspace parsing / review composition use the stdio transport described above.

VS Code .vscode/mcp.json:

{
  "servers": {
    "cpsagentkit-hosted": {
      "type": "http",
      "url": "https://ca-mcp-6frsoq6bw5vuk.salmonstone-8b839f60.uksouth.azurecontainerapps.io/mcp"
    }
  }
}

Deployment details, including how to enable optional API-key auth, live in docs/deployment/mcp-server-azure.md.