clark-mcp (archived)

This repository is archived. The MCP server moved into the main Clark repo. See clark/mcp/ for the code and README.md → MCP integration for setup with Claude Desktop, Cursor, Continue, Zed, and other MCP-aware hosts.

TL;DR migration

Old:

git clone https://github.com/jarmstrong158/clark-mcp
pip install -e clark-mcp
clark-mcp                          # MCP stdio entrypoint

New (single repo):

git clone https://github.com/jarmstrong158/Clark
pip install -e "Clark[mcp,serve]"
clark serve --model <ckpt> --facilities-dir clark/data/configs --port 8000   # one shell
clark mcp                                                                     # another shell, or wired into your MCP host

MCP host config (Claude Desktop / Cursor / etc.):

{
  "mcpServers": {
    "clark": {
      "command": "clark",
      "args": ["mcp"],
      "env": {"CLARK_API_URL": "http://127.0.0.1:8000"}
    }
  }
}

Related MCP server: KPI-Lens

Why this repo was retired

clark-mcp was built around a local-LLM-driven natural-language UX (Hermes-3-8B in Ollama, QLoRA fine-tune, schema-constrained decoding, a custom chat web UI). When Clark grew an operations dashboard (clark ops) that covers the same operator workflows with forms over the trained policy, the natural-language layer became redundant for the headline use case and the maintenance cost stopped being worth it.

What survived the cut: the MCP server itself, since any MCP-aware host can still drive Clark in plain English using that host's own model (no fine-tune required, no Ollama, no extra weights to ship). That code, roughly 200 lines, is the right size to live as a subpackage of Clark rather than its own repo with its own license, CI, and release cadence.

What's still here for the record

• Source frozen at the last working state of the local-LLM path: clark_mcp/agent.py (Hermes-3 client), clark_mcp/web/ (chat UI), clark_mcp/chat_format.py (tool-calling protocol), clark_mcp/briefing.py (morning briefing generator), and the finetune/ directory with the dataset, eval.py, and PHASE3.md scorecard. None of it is maintained.

• Git history is intact: the fine-tune dataset construction notes, the held-out eval methodology, and the tools/live_audit.py autoregressive probe suite are still browsable as a record of how a small local model was taught to use a domain API truthfully.

• No releases. The pyproject.toml script entries (clark-mcp, clark-mcp-web, clark-mcp-briefing) still install but target a localhost clark serve that no longer pairs naturally with this repo's tooling. Use the new home above.

License

PolyForm Noncommercial 1.0.0, same as Clark. Read, study, fork; commercial use of Clark or its derivatives requires a separate agreement (see Clark's repo).