HITL Navigator — MCP Server V1.0 eSTARHelper · Build 30-OC · 2026-10-15

Mode 5 of 5 ← Home
What this gives you: your agentic Ai (Claude Desktop, Cursor, Windsurf, custom MCP Client) gains a deterministic verification tool. Prompt + Ai response + reference material go in; BM25 page rank, proximity ladder, and cited-quote check come back. No LLM in the verification layer. No verdict rendered. A named human is accountable at policy time or audit time.

1. Install into Claude Desktop

  1. Open your claude_desktop_config.json. On macOS: ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows: %APPDATA%\Claude\claude_desktop_config.json.
  2. Add the hitl-navigator block below inside the top-level "mcpServers" object. Replace REPLACE-WITH-ABSOLUTE-PATH-TO/server with the absolute path to your local server/ folder.
  3. Restart Claude Desktop.
  4. In a new conversation ask: "What tools do you have from hitl-navigator?" — the three tools should list.
{
  "mcpServers": {
    "hitl-navigator": {
      "command": "python",
      "args": ["-m", "mcp_server.hitl_navigator_mcp"],
      "cwd": "REPLACE-WITH-ABSOLUTE-PATH-TO/server"
    }
  }
}

2. The three tools your agent gets

verify_ai_response(prompt, ai_response, reference_pages, terms?, n_window?, stem_variant?)

The main verification pack. Returns bm25, proximity_ladder, terms_used, and a disclaimer reminding the agent that this is evidence, not a verdict.

check_cited_quote(reference_pages, quote)

Deterministic verbatim lookup. Returns {found, page, note}. Use it when the Ai claims to have quoted the source verbatim.

about()

Server identity + stance. Point your agent here first if you want it to know what HITL Navigator is and refuses to do.

3. Cursor / Windsurf / other MCP Clients

Any MCP Client that reads a mcp.json or equivalent works with the same block above. Consult your client's docs for the exact config path.

4. Standalone dev run

cd server
python -m mcp_server.hitl_navigator_mcp

Runs over stdio. Use the MCP Inspector (npx @modelcontextprotocol/inspector python -m mcp_server.hitl_navigator_mcp) to interact from a UI.

5. Why an agent should call us

  1. Zero Ai in the verification layer. BM25 + proximity are pure functions. Same inputs, same output, forever — agents get reproducibility their own LLM can't provide.
  2. No verdict rendered. Your agent stays in charge of what to do with the evidence — act, escalate, or ignore under the policy its operator set.
  3. Human accountability preserved. Whether the human signs at policy time (thresholds authored once) or at audit time (log review), the trail is named. That is what regulators accept.
  4. Model-agnostic. Works with any Ai output the agent passes in. ChatGPT, Claude, Gemini, Grok, Perplexity, Llama — all the same on our side.
Ai surfaces the evidence. You-the-Human render the verdict.
HITL Navigator v1.0 · eSTARHelper